1 //! # Rust Compiler Self-Profiling
3 //! This module implements the basic framework for the compiler's self-
4 //! profiling support. It provides the `SelfProfiler` type which enables
5 //! recording "events". An event is something that starts and ends at a given
6 //! point in time and has an ID and a kind attached to it. This allows for
7 //! tracing the compiler's activity.
9 //! Internally this module uses the custom tailored [measureme][mm] crate for
10 //! efficiently recording events to disk in a compact format that can be
11 //! post-processed and analyzed by the suite of tools in the `measureme`
12 //! project. The highest priority for the tracing framework is on incurring as
13 //! little overhead as possible.
18 //! Events have a few properties:
20 //! - The `event_kind` designates the broad category of an event (e.g. does it
21 //! correspond to the execution of a query provider or to loading something
22 //! from the incr. comp. on-disk cache, etc).
23 //! - The `event_id` designates the query invocation or function call it
24 //! corresponds to, possibly including the query key or function arguments.
25 //! - Each event stores the ID of the thread it was recorded on.
26 //! - The timestamp stores beginning and end of the event, or the single point
27 //! in time it occurred at for "instant" events.
30 //! ## Event Filtering
32 //! Event generation can be filtered by event kind. Recording all possible
33 //! events generates a lot of data, much of which is not needed for most kinds
34 //! of analysis. So, in order to keep overhead as low as possible for a given
35 //! use case, the `SelfProfiler` will only record the kinds of events that
36 //! pass the filter specified as a command line argument to the compiler.
39 //! ## `event_id` Assignment
41 //! As far as `measureme` is concerned, `event_id`s are just strings. However,
42 //! it would incur too much overhead to generate and persist each `event_id`
43 //! string at the point where the event is recorded. In order to make this more
44 //! efficient `measureme` has two features:
46 //! - Strings can share their content, so that re-occurring parts don't have to
47 //! be copied over and over again. One allocates a string in `measureme` and
48 //! gets back a `StringId`. This `StringId` is then used to refer to that
49 //! string. `measureme` strings are actually DAGs of string components so that
50 //! arbitrary sharing of substrings can be done efficiently. This is useful
51 //! because `event_id`s contain lots of redundant text like query names or
52 //! def-path components.
54 //! - `StringId`s can be "virtual" which means that the client picks a numeric
55 //! ID according to some application-specific scheme and can later make that
56 //! ID be mapped to an actual string. This is used to cheaply generate
57 //! `event_id`s while the events actually occur, causing little timing
58 //! distortion, and then later map those `StringId`s, in bulk, to actual
59 //! `event_id` strings. This way the largest part of the tracing overhead is
60 //! localized to one contiguous chunk of time.
62 //! How are these `event_id`s generated in the compiler? For things that occur
63 //! infrequently (e.g. "generic activities"), we just allocate the string the
64 //! first time it is used and then keep the `StringId` in a hash table. This
65 //! is implemented in `SelfProfiler::get_or_alloc_cached_string()`.
67 //! For queries it gets more interesting: First we need a unique numeric ID for
68 //! each query invocation (the `QueryInvocationId`). This ID is used as the
69 //! virtual `StringId` we use as `event_id` for a given event. This ID has to
70 //! be available both when the query is executed and later, together with the
71 //! query key, when we allocate the actual `event_id` strings in bulk.
73 //! We could make the compiler generate and keep track of such an ID for each
74 //! query invocation but luckily we already have something that fits all the
75 //! the requirements: the query's `DepNodeIndex`. So we use the numeric value
76 //! of the `DepNodeIndex` as `event_id` when recording the event and then,
77 //! just before the query context is dropped, we walk the entire query cache
78 //! (which stores the `DepNodeIndex` along with the query key for each
79 //! invocation) and allocate the corresponding strings together with a mapping
80 //! for `DepNodeIndex as StringId`.
82 //! [mm]: https://github.com/rust-lang/measureme/
85 use crate::fx::FxHashMap;
87 use std::borrow::Borrow;
88 use std::collections::hash_map::Entry;
89 use std::convert::Into;
90 use std::error::Error;
95 use std::time::{Duration, Instant};
97 pub use measureme::EventId;
98 use measureme::{EventIdBuilder, Profiler, SerializableString, StringId};
99 use parking_lot::RwLock;
100 use smallvec::SmallVec;
102 bitflags::bitflags! {
103 struct EventFilter: u32 {
104 const GENERIC_ACTIVITIES = 1 << 0;
105 const QUERY_PROVIDERS = 1 << 1;
106 const QUERY_CACHE_HITS = 1 << 2;
107 const QUERY_BLOCKED = 1 << 3;
108 const INCR_CACHE_LOADS = 1 << 4;
110 const QUERY_KEYS = 1 << 5;
111 const FUNCTION_ARGS = 1 << 6;
113 const INCR_RESULT_HASHING = 1 << 8;
114 const ARTIFACT_SIZES = 1 << 9;
116 const DEFAULT = Self::GENERIC_ACTIVITIES.bits |
117 Self::QUERY_PROVIDERS.bits |
118 Self::QUERY_BLOCKED.bits |
119 Self::INCR_CACHE_LOADS.bits |
120 Self::INCR_RESULT_HASHING.bits |
121 Self::ARTIFACT_SIZES.bits;
123 const ARGS = Self::QUERY_KEYS.bits | Self::FUNCTION_ARGS.bits;
127 // keep this in sync with the `-Z self-profile-events` help message in rustc_session/options.rs
128 const EVENT_FILTERS_BY_NAME: &[(&str, EventFilter)] = &[
129 ("none", EventFilter::empty()),
130 ("all", EventFilter::all()),
131 ("default", EventFilter::DEFAULT),
132 ("generic-activity", EventFilter::GENERIC_ACTIVITIES),
133 ("query-provider", EventFilter::QUERY_PROVIDERS),
134 ("query-cache-hit", EventFilter::QUERY_CACHE_HITS),
135 ("query-blocked", EventFilter::QUERY_BLOCKED),
136 ("incr-cache-load", EventFilter::INCR_CACHE_LOADS),
137 ("query-keys", EventFilter::QUERY_KEYS),
138 ("function-args", EventFilter::FUNCTION_ARGS),
139 ("args", EventFilter::ARGS),
140 ("llvm", EventFilter::LLVM),
141 ("incr-result-hashing", EventFilter::INCR_RESULT_HASHING),
142 ("artifact-sizes", EventFilter::ARTIFACT_SIZES),
145 /// Something that uniquely identifies a query invocation.
146 pub struct QueryInvocationId(pub u32);
148 /// A reference to the SelfProfiler. It can be cloned and sent across thread
149 /// boundaries at will.
151 pub struct SelfProfilerRef {
152 // This field is `None` if self-profiling is disabled for the current
153 // compilation session.
154 profiler: Option<Arc<SelfProfiler>>,
156 // We store the filter mask directly in the reference because that doesn't
157 // cost anything and allows for filtering with checking if the profiler is
159 event_filter_mask: EventFilter,
161 // Print verbose generic activities to stdout
162 print_verbose_generic_activities: bool,
164 // Print extra verbose generic activities to stdout
165 print_extra_verbose_generic_activities: bool,
168 impl SelfProfilerRef {
170 profiler: Option<Arc<SelfProfiler>>,
171 print_verbose_generic_activities: bool,
172 print_extra_verbose_generic_activities: bool,
173 ) -> SelfProfilerRef {
174 // If there is no SelfProfiler then the filter mask is set to NONE,
175 // ensuring that nothing ever tries to actually access it.
176 let event_filter_mask =
177 profiler.as_ref().map_or(EventFilter::empty(), |p| p.event_filter_mask);
182 print_verbose_generic_activities,
183 print_extra_verbose_generic_activities,
187 /// This shim makes sure that calls only get executed if the filter mask
188 /// lets them pass. It also contains some trickery to make sure that
189 /// code is optimized for non-profiling compilation sessions, i.e. anything
190 /// past the filter check is never inlined so it doesn't clutter the fast
193 fn exec<F>(&self, event_filter: EventFilter, f: F) -> TimingGuard<'_>
195 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
198 fn cold_call<F>(profiler_ref: &SelfProfilerRef, f: F) -> TimingGuard<'_>
200 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
202 let profiler = profiler_ref.profiler.as_ref().unwrap();
206 if unlikely!(self.event_filter_mask.contains(event_filter)) {
213 /// Start profiling a verbose generic activity. Profiling continues until the
214 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
215 /// a measureme event, "verbose" generic activities also print a timing entry to
216 /// stdout if the compiler is invoked with -Ztime or -Ztime-passes.
217 pub fn verbose_generic_activity<'a>(
219 event_label: &'static str,
220 ) -> VerboseTimingGuard<'a> {
222 if self.print_verbose_generic_activities { Some(event_label.to_owned()) } else { None };
224 VerboseTimingGuard::start(message, self.generic_activity(event_label))
227 /// Start profiling an extra verbose generic activity. Profiling continues until the
228 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
229 /// a measureme event, "extra verbose" generic activities also print a timing entry to
230 /// stdout if the compiler is invoked with -Ztime-passes.
231 pub fn extra_verbose_generic_activity<'a, A>(
233 event_label: &'static str,
235 ) -> VerboseTimingGuard<'a>
237 A: Borrow<str> + Into<String>,
239 let message = if self.print_extra_verbose_generic_activities {
240 Some(format!("{}({})", event_label, event_arg.borrow()))
245 VerboseTimingGuard::start(message, self.generic_activity_with_arg(event_label, event_arg))
248 /// Start profiling a generic activity. Profiling continues until the
249 /// TimingGuard returned from this call is dropped.
251 pub fn generic_activity(&self, event_label: &'static str) -> TimingGuard<'_> {
252 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
253 let event_label = profiler.get_or_alloc_cached_string(event_label);
254 let event_id = EventId::from_label(event_label);
255 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
259 /// Start profiling with some event filter for a given event. Profiling continues until the
260 /// TimingGuard returned from this call is dropped.
262 pub fn generic_activity_with_event_id(&self, event_id: EventId) -> TimingGuard<'_> {
263 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
264 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
268 /// Start profiling a generic activity. Profiling continues until the
269 /// TimingGuard returned from this call is dropped.
271 pub fn generic_activity_with_arg<A>(
273 event_label: &'static str,
277 A: Borrow<str> + Into<String>,
279 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
280 let builder = EventIdBuilder::new(&profiler.profiler);
281 let event_label = profiler.get_or_alloc_cached_string(event_label);
282 let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) {
283 let event_arg = profiler.get_or_alloc_cached_string(event_arg);
284 builder.from_label_and_arg(event_label, event_arg)
286 builder.from_label(event_label)
288 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
292 /// Start profiling a generic activity, allowing costly arguments to be recorded. Profiling
293 /// continues until the `TimingGuard` returned from this call is dropped.
295 /// If the arguments to a generic activity are cheap to create, use `generic_activity_with_arg`
296 /// or `generic_activity_with_args` for their simpler API. However, if they are costly or
297 /// require allocation in sufficiently hot contexts, then this allows for a closure to be called
298 /// only when arguments were asked to be recorded via `-Z self-profile-events=args`.
300 /// In this case, the closure will be passed a `&mut EventArgRecorder`, to help with recording
301 /// one or many arguments within the generic activity being profiled, by calling its
302 /// `record_arg` method for example.
304 /// This `EventArgRecorder` may implement more specific traits from other rustc crates, e.g. for
305 /// richer handling of rustc-specific argument types, while keeping this single entry-point API
306 /// for recording arguments.
308 /// Note: recording at least one argument is *required* for the self-profiler to create the
309 /// `TimingGuard`. A panic will be triggered if that doesn't happen. This function exists
310 /// explicitly to record arguments, so it fails loudly when there are none to record.
313 pub fn generic_activity_with_arg_recorder<F>(
315 event_label: &'static str,
319 F: FnMut(&mut EventArgRecorder<'_>),
321 // Ensure this event will only be recorded when self-profiling is turned on.
322 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
323 let builder = EventIdBuilder::new(&profiler.profiler);
324 let event_label = profiler.get_or_alloc_cached_string(event_label);
326 // Ensure the closure to create event arguments will only be called when argument
327 // recording is turned on.
328 let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) {
329 // Set up the builder and call the user-provided closure to record potentially
330 // costly event arguments.
331 let mut recorder = EventArgRecorder { profiler, args: SmallVec::new() };
334 // It is expected that the closure will record at least one argument. If that
335 // doesn't happen, it's a bug: we've been explicitly called in order to record
336 // arguments, so we fail loudly when there are none to record.
337 if recorder.args.is_empty() {
339 "The closure passed to `generic_activity_with_arg_recorder` needs to \
340 record at least one argument"
344 builder.from_label_and_args(event_label, &recorder.args)
346 builder.from_label(event_label)
348 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
352 /// Record the size of an artifact that the compiler produces
354 /// `artifact_kind` is the class of artifact (e.g., query_cache, object_file, etc.)
355 /// `artifact_name` is an identifier to the specific artifact being stored (usually a filename)
357 pub fn artifact_size<A>(&self, artifact_kind: &str, artifact_name: A, size: u64)
359 A: Borrow<str> + Into<String>,
361 drop(self.exec(EventFilter::ARTIFACT_SIZES, |profiler| {
362 let builder = EventIdBuilder::new(&profiler.profiler);
363 let event_label = profiler.get_or_alloc_cached_string(artifact_kind);
364 let event_arg = profiler.get_or_alloc_cached_string(artifact_name);
365 let event_id = builder.from_label_and_arg(event_label, event_arg);
366 let thread_id = get_thread_id();
368 profiler.profiler.record_integer_event(
369 profiler.artifact_size_event_kind,
380 pub fn generic_activity_with_args(
382 event_label: &'static str,
383 event_args: &[String],
384 ) -> TimingGuard<'_> {
385 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
386 let builder = EventIdBuilder::new(&profiler.profiler);
387 let event_label = profiler.get_or_alloc_cached_string(event_label);
388 let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) {
389 let event_args: Vec<_> = event_args
391 .map(|s| profiler.get_or_alloc_cached_string(&s[..]))
393 builder.from_label_and_args(event_label, &event_args)
395 builder.from_label(event_label)
397 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
401 /// Start profiling a query provider. Profiling continues until the
402 /// TimingGuard returned from this call is dropped.
404 pub fn query_provider(&self) -> TimingGuard<'_> {
405 self.exec(EventFilter::QUERY_PROVIDERS, |profiler| {
406 TimingGuard::start(profiler, profiler.query_event_kind, EventId::INVALID)
410 /// Record a query in-memory cache hit.
412 pub fn query_cache_hit(&self, query_invocation_id: QueryInvocationId) {
413 self.instant_query_event(
414 |profiler| profiler.query_cache_hit_event_kind,
416 EventFilter::QUERY_CACHE_HITS,
420 /// Start profiling a query being blocked on a concurrent execution.
421 /// Profiling continues until the TimingGuard returned from this call is
424 pub fn query_blocked(&self) -> TimingGuard<'_> {
425 self.exec(EventFilter::QUERY_BLOCKED, |profiler| {
426 TimingGuard::start(profiler, profiler.query_blocked_event_kind, EventId::INVALID)
430 /// Start profiling how long it takes to load a query result from the
431 /// incremental compilation on-disk cache. Profiling continues until the
432 /// TimingGuard returned from this call is dropped.
434 pub fn incr_cache_loading(&self) -> TimingGuard<'_> {
435 self.exec(EventFilter::INCR_CACHE_LOADS, |profiler| {
438 profiler.incremental_load_result_event_kind,
444 /// Start profiling how long it takes to hash query results for incremental compilation.
445 /// Profiling continues until the TimingGuard returned from this call is dropped.
447 pub fn incr_result_hashing(&self) -> TimingGuard<'_> {
448 self.exec(EventFilter::INCR_RESULT_HASHING, |profiler| {
451 profiler.incremental_result_hashing_event_kind,
458 fn instant_query_event(
460 event_kind: fn(&SelfProfiler) -> StringId,
461 query_invocation_id: QueryInvocationId,
462 event_filter: EventFilter,
464 drop(self.exec(event_filter, |profiler| {
465 let event_id = StringId::new_virtual(query_invocation_id.0);
466 let thread_id = get_thread_id();
468 profiler.profiler.record_instant_event(
469 event_kind(profiler),
470 EventId::from_virtual(event_id),
478 pub fn with_profiler(&self, f: impl FnOnce(&SelfProfiler)) {
479 if let Some(profiler) = &self.profiler {
484 /// Gets a `StringId` for the given string. This method makes sure that
485 /// any strings going through it will only be allocated once in the
487 /// Returns `None` if the self-profiling is not enabled.
488 pub fn get_or_alloc_cached_string(&self, s: &str) -> Option<StringId> {
489 self.profiler.as_ref().map(|p| p.get_or_alloc_cached_string(s))
493 pub fn enabled(&self) -> bool {
494 self.profiler.is_some()
498 pub fn llvm_recording_enabled(&self) -> bool {
499 self.event_filter_mask.contains(EventFilter::LLVM)
502 pub fn get_self_profiler(&self) -> Option<Arc<SelfProfiler>> {
503 self.profiler.clone()
507 /// A helper for recording costly arguments to self-profiling events. Used with
508 /// `SelfProfilerRef::generic_activity_with_arg_recorder`.
509 pub struct EventArgRecorder<'p> {
510 /// The `SelfProfiler` used to intern the event arguments that users will ask to record.
511 profiler: &'p SelfProfiler,
513 /// The interned event arguments to be recorded in the generic activity event.
515 /// The most common case, when actually recording event arguments, is to have one argument. Then
516 /// followed by recording two, in a couple places.
517 args: SmallVec<[StringId; 2]>,
520 impl EventArgRecorder<'_> {
521 /// Records a single argument within the current generic activity being profiled.
523 /// Note: when self-profiling with costly event arguments, at least one argument
524 /// needs to be recorded. A panic will be triggered if that doesn't happen.
525 pub fn record_arg<A>(&mut self, event_arg: A)
527 A: Borrow<str> + Into<String>,
529 let event_arg = self.profiler.get_or_alloc_cached_string(event_arg);
530 self.args.push(event_arg);
534 pub struct SelfProfiler {
536 event_filter_mask: EventFilter,
538 string_cache: RwLock<FxHashMap<String, StringId>>,
540 query_event_kind: StringId,
541 generic_activity_event_kind: StringId,
542 incremental_load_result_event_kind: StringId,
543 incremental_result_hashing_event_kind: StringId,
544 query_blocked_event_kind: StringId,
545 query_cache_hit_event_kind: StringId,
546 artifact_size_event_kind: StringId,
551 output_directory: &Path,
552 crate_name: Option<&str>,
553 event_filters: &Option<Vec<String>>,
554 ) -> Result<SelfProfiler, Box<dyn Error + Send + Sync>> {
555 fs::create_dir_all(output_directory)?;
557 let crate_name = crate_name.unwrap_or("unknown-crate");
558 let filename = format!("{}-{}.rustc_profile", crate_name, process::id());
559 let path = output_directory.join(&filename);
560 let profiler = Profiler::new(&path)?;
562 let query_event_kind = profiler.alloc_string("Query");
563 let generic_activity_event_kind = profiler.alloc_string("GenericActivity");
564 let incremental_load_result_event_kind = profiler.alloc_string("IncrementalLoadResult");
565 let incremental_result_hashing_event_kind =
566 profiler.alloc_string("IncrementalResultHashing");
567 let query_blocked_event_kind = profiler.alloc_string("QueryBlocked");
568 let query_cache_hit_event_kind = profiler.alloc_string("QueryCacheHit");
569 let artifact_size_event_kind = profiler.alloc_string("ArtifactSize");
571 let mut event_filter_mask = EventFilter::empty();
573 if let Some(ref event_filters) = *event_filters {
574 let mut unknown_events = vec![];
575 for item in event_filters {
576 if let Some(&(_, mask)) =
577 EVENT_FILTERS_BY_NAME.iter().find(|&(name, _)| name == item)
579 event_filter_mask |= mask;
581 unknown_events.push(item.clone());
585 // Warn about any unknown event names
586 if !unknown_events.is_empty() {
587 unknown_events.sort();
588 unknown_events.dedup();
591 "Unknown self-profiler events specified: {}. Available options are: {}.",
592 unknown_events.join(", "),
593 EVENT_FILTERS_BY_NAME
595 .map(|&(name, _)| name.to_string())
601 event_filter_mask = EventFilter::DEFAULT;
607 string_cache: RwLock::new(FxHashMap::default()),
609 generic_activity_event_kind,
610 incremental_load_result_event_kind,
611 incremental_result_hashing_event_kind,
612 query_blocked_event_kind,
613 query_cache_hit_event_kind,
614 artifact_size_event_kind,
618 /// Allocates a new string in the profiling data. Does not do any caching
619 /// or deduplication.
620 pub fn alloc_string<STR: SerializableString + ?Sized>(&self, s: &STR) -> StringId {
621 self.profiler.alloc_string(s)
624 /// Gets a `StringId` for the given string. This method makes sure that
625 /// any strings going through it will only be allocated once in the
627 pub fn get_or_alloc_cached_string<A>(&self, s: A) -> StringId
629 A: Borrow<str> + Into<String>,
631 // Only acquire a read-lock first since we assume that the string is
632 // already present in the common case.
634 let string_cache = self.string_cache.read();
636 if let Some(&id) = string_cache.get(s.borrow()) {
641 let mut string_cache = self.string_cache.write();
642 // Check if the string has already been added in the small time window
643 // between dropping the read lock and acquiring the write lock.
644 match string_cache.entry(s.into()) {
645 Entry::Occupied(e) => *e.get(),
646 Entry::Vacant(e) => {
647 let string_id = self.profiler.alloc_string(&e.key()[..]);
653 pub fn map_query_invocation_id_to_string(&self, from: QueryInvocationId, to: StringId) {
654 let from = StringId::new_virtual(from.0);
655 self.profiler.map_virtual_to_concrete_string(from, to);
658 pub fn bulk_map_query_invocation_id_to_single_string<I>(&self, from: I, to: StringId)
660 I: Iterator<Item = QueryInvocationId> + ExactSizeIterator,
662 let from = from.map(|qid| StringId::new_virtual(qid.0));
663 self.profiler.bulk_map_virtual_to_single_concrete_string(from, to);
666 pub fn query_key_recording_enabled(&self) -> bool {
667 self.event_filter_mask.contains(EventFilter::QUERY_KEYS)
670 pub fn event_id_builder(&self) -> EventIdBuilder<'_> {
671 EventIdBuilder::new(&self.profiler)
676 pub struct TimingGuard<'a>(Option<measureme::TimingGuard<'a>>);
678 impl<'a> TimingGuard<'a> {
681 profiler: &'a SelfProfiler,
682 event_kind: StringId,
684 ) -> TimingGuard<'a> {
685 let thread_id = get_thread_id();
686 let raw_profiler = &profiler.profiler;
688 raw_profiler.start_recording_interval_event(event_kind, event_id, thread_id);
689 TimingGuard(Some(timing_guard))
693 pub fn finish_with_query_invocation_id(self, query_invocation_id: QueryInvocationId) {
694 if let Some(guard) = self.0 {
696 let event_id = StringId::new_virtual(query_invocation_id.0);
697 let event_id = EventId::from_virtual(event_id);
698 guard.finish_with_override_event_id(event_id);
704 pub fn none() -> TimingGuard<'a> {
709 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
716 pub struct VerboseTimingGuard<'a> {
717 start_and_message: Option<(Instant, Option<usize>, String)>,
718 _guard: TimingGuard<'a>,
721 impl<'a> VerboseTimingGuard<'a> {
722 pub fn start(message: Option<String>, _guard: TimingGuard<'a>) -> Self {
725 start_and_message: message.map(|msg| (Instant::now(), get_resident_set_size(), msg)),
730 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
736 impl Drop for VerboseTimingGuard<'_> {
738 if let Some((start_time, start_rss, ref message)) = self.start_and_message {
739 let end_rss = get_resident_set_size();
740 let dur = start_time.elapsed();
742 if should_print_passes(dur, start_rss, end_rss) {
743 print_time_passes_entry(&message, dur, start_rss, end_rss);
749 fn should_print_passes(dur: Duration, start_rss: Option<usize>, end_rss: Option<usize>) -> bool {
750 if dur.as_millis() > 5 {
754 if let (Some(start_rss), Some(end_rss)) = (start_rss, end_rss) {
755 let change_rss = end_rss.abs_diff(start_rss);
764 pub fn print_time_passes_entry(
767 start_rss: Option<usize>,
768 end_rss: Option<usize>,
770 let rss_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as usize;
771 let rss_change_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as i128;
773 let mem_string = match (start_rss, end_rss) {
774 (Some(start_rss), Some(end_rss)) => {
775 let change_rss = end_rss as i128 - start_rss as i128;
778 "; rss: {:>4}MB -> {:>4}MB ({:>+5}MB)",
779 rss_to_mb(start_rss),
781 rss_change_to_mb(change_rss),
784 (Some(start_rss), None) => format!("; rss start: {:>4}MB", rss_to_mb(start_rss)),
785 (None, Some(end_rss)) => format!("; rss end: {:>4}MB", rss_to_mb(end_rss)),
786 (None, None) => String::new(),
789 eprintln!("time: {:>7}{}\t{}", duration_to_secs_str(dur), mem_string, what);
792 // Hack up our own formatting for the duration to make it easier for scripts
793 // to parse (always use the same number of decimal places and the same unit).
794 pub fn duration_to_secs_str(dur: std::time::Duration) -> String {
795 format!("{:.3}", dur.as_secs_f64())
798 fn get_thread_id() -> u32 {
799 std::thread::current().id().as_u64().get() as u32
805 pub fn get_resident_set_size() -> Option<usize> {
806 use std::mem::{self, MaybeUninit};
807 use winapi::shared::minwindef::DWORD;
808 use winapi::um::processthreadsapi::GetCurrentProcess;
809 use winapi::um::psapi::{GetProcessMemoryInfo, PROCESS_MEMORY_COUNTERS};
811 let mut pmc = MaybeUninit::<PROCESS_MEMORY_COUNTERS>::uninit();
813 GetProcessMemoryInfo(GetCurrentProcess(), pmc.as_mut_ptr(), mem::size_of_val(&pmc) as DWORD)
817 let pmc = unsafe { pmc.assume_init() };
818 Some(pmc.WorkingSetSize as usize)
822 } else if #[cfg(unix)] {
823 pub fn get_resident_set_size() -> Option<usize> {
825 let contents = fs::read("/proc/self/statm").ok()?;
826 let contents = String::from_utf8(contents).ok()?;
827 let s = contents.split_whitespace().nth(field)?;
828 let npages = s.parse::<usize>().ok()?;
832 pub fn get_resident_set_size() -> Option<usize> {