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/
84 use crate::fx::FxHashMap;
86 use std::error::Error;
91 use std::thread::ThreadId;
92 use std::time::{Duration, Instant};
95 use measureme::{EventId, EventIdBuilder, SerializableString, StringId};
96 use parking_lot::RwLock;
98 /// MmapSerializatioSink is faster on macOS and Linux
99 /// but FileSerializationSink is faster on Windows
101 type SerializationSink = measureme::MmapSerializationSink;
103 type SerializationSink = measureme::FileSerializationSink;
105 type Profiler = measureme::Profiler<SerializationSink>;
107 #[derive(Clone, Copy, Debug, PartialEq, Eq, Ord, PartialOrd)]
108 pub enum ProfileCategory {
118 bitflags::bitflags! {
119 struct EventFilter: u32 {
120 const GENERIC_ACTIVITIES = 1 << 0;
121 const QUERY_PROVIDERS = 1 << 1;
122 const QUERY_CACHE_HITS = 1 << 2;
123 const QUERY_BLOCKED = 1 << 3;
124 const INCR_CACHE_LOADS = 1 << 4;
126 const QUERY_KEYS = 1 << 5;
128 const DEFAULT = Self::GENERIC_ACTIVITIES.bits |
129 Self::QUERY_PROVIDERS.bits |
130 Self::QUERY_BLOCKED.bits |
131 Self::INCR_CACHE_LOADS.bits;
133 // empty() and none() aren't const-fns unfortunately
135 const ALL = !Self::NONE.bits;
139 // keep this in sync with the `-Z self-profile-events` help message in librustc_session/options.rs
140 const EVENT_FILTERS_BY_NAME: &[(&str, EventFilter)] = &[
141 ("none", EventFilter::NONE),
142 ("all", EventFilter::ALL),
143 ("default", EventFilter::DEFAULT),
144 ("generic-activity", EventFilter::GENERIC_ACTIVITIES),
145 ("query-provider", EventFilter::QUERY_PROVIDERS),
146 ("query-cache-hit", EventFilter::QUERY_CACHE_HITS),
147 ("query-blocked", EventFilter::QUERY_BLOCKED),
148 ("incr-cache-load", EventFilter::INCR_CACHE_LOADS),
149 ("query-keys", EventFilter::QUERY_KEYS),
152 fn thread_id_to_u32(tid: ThreadId) -> u32 {
153 unsafe { std::mem::transmute::<ThreadId, u64>(tid) as u32 }
156 /// Something that uniquely identifies a query invocation.
157 pub struct QueryInvocationId(pub u32);
159 /// A reference to the SelfProfiler. It can be cloned and sent across thread
160 /// boundaries at will.
162 pub struct SelfProfilerRef {
163 // This field is `None` if self-profiling is disabled for the current
164 // compilation session.
165 profiler: Option<Arc<SelfProfiler>>,
167 // We store the filter mask directly in the reference because that doesn't
168 // cost anything and allows for filtering with checking if the profiler is
170 event_filter_mask: EventFilter,
172 // Print verbose generic activities to stdout
173 print_verbose_generic_activities: bool,
175 // Print extra verbose generic activities to stdout
176 print_extra_verbose_generic_activities: bool,
179 impl SelfProfilerRef {
181 profiler: Option<Arc<SelfProfiler>>,
182 print_verbose_generic_activities: bool,
183 print_extra_verbose_generic_activities: bool,
184 ) -> SelfProfilerRef {
185 // If there is no SelfProfiler then the filter mask is set to NONE,
186 // ensuring that nothing ever tries to actually access it.
187 let event_filter_mask =
188 profiler.as_ref().map(|p| p.event_filter_mask).unwrap_or(EventFilter::NONE);
193 print_verbose_generic_activities,
194 print_extra_verbose_generic_activities,
198 // This shim makes sure that calls only get executed if the filter mask
199 // lets them pass. It also contains some trickery to make sure that
200 // code is optimized for non-profiling compilation sessions, i.e. anything
201 // past the filter check is never inlined so it doesn't clutter the fast
204 fn exec<F>(&self, event_filter: EventFilter, f: F) -> TimingGuard<'_>
206 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
209 fn cold_call<F>(profiler_ref: &SelfProfilerRef, f: F) -> TimingGuard<'_>
211 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
213 let profiler = profiler_ref.profiler.as_ref().unwrap();
217 if unlikely!(self.event_filter_mask.contains(event_filter)) {
224 /// Start profiling a verbose generic activity. Profiling continues until the
225 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
226 /// a measureme event, "verbose" generic activities also print a timing entry to
227 /// stdout if the compiler is invoked with -Ztime or -Ztime-passes.
229 pub fn verbose_generic_activity<'a>(
231 event_id: &'static str,
232 ) -> VerboseTimingGuard<'a> {
233 VerboseTimingGuard::start(
235 self.print_verbose_generic_activities,
236 self.generic_activity(event_id),
240 /// Start profiling a extra verbose generic activity. Profiling continues until the
241 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
242 /// a measureme event, "extra verbose" generic activities also print a timing entry to
243 /// stdout if the compiler is invoked with -Ztime-passes.
245 pub fn extra_verbose_generic_activity<'a>(
248 ) -> VerboseTimingGuard<'a> {
249 // FIXME: This does not yet emit a measureme event
250 // because callers encode arguments into `event_id`.
251 VerboseTimingGuard::start(
253 self.print_extra_verbose_generic_activities,
258 /// Start profiling a generic activity. Profiling continues until the
259 /// TimingGuard returned from this call is dropped.
261 pub fn generic_activity(&self, event_id: &'static str) -> TimingGuard<'_> {
262 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
263 let event_id = profiler.get_or_alloc_cached_string(event_id);
264 let event_id = EventId::from_label(event_id);
265 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
269 /// Start profiling a query provider. Profiling continues until the
270 /// TimingGuard returned from this call is dropped.
272 pub fn query_provider(&self) -> TimingGuard<'_> {
273 self.exec(EventFilter::QUERY_PROVIDERS, |profiler| {
274 TimingGuard::start(profiler, profiler.query_event_kind, EventId::INVALID)
278 /// Record a query in-memory cache hit.
280 pub fn query_cache_hit(&self, query_invocation_id: QueryInvocationId) {
281 self.instant_query_event(
282 |profiler| profiler.query_cache_hit_event_kind,
284 EventFilter::QUERY_CACHE_HITS,
288 /// Start profiling a query being blocked on a concurrent execution.
289 /// Profiling continues until the TimingGuard returned from this call is
292 pub fn query_blocked(&self) -> TimingGuard<'_> {
293 self.exec(EventFilter::QUERY_BLOCKED, |profiler| {
294 TimingGuard::start(profiler, profiler.query_blocked_event_kind, EventId::INVALID)
298 /// Start profiling how long it takes to load a query result from the
299 /// incremental compilation on-disk cache. Profiling continues until the
300 /// TimingGuard returned from this call is dropped.
302 pub fn incr_cache_loading(&self) -> TimingGuard<'_> {
303 self.exec(EventFilter::INCR_CACHE_LOADS, |profiler| {
306 profiler.incremental_load_result_event_kind,
313 fn instant_query_event(
315 event_kind: fn(&SelfProfiler) -> StringId,
316 query_invocation_id: QueryInvocationId,
317 event_filter: EventFilter,
319 drop(self.exec(event_filter, |profiler| {
320 let event_id = StringId::new_virtual(query_invocation_id.0);
321 let thread_id = thread_id_to_u32(std::thread::current().id());
323 profiler.profiler.record_instant_event(
324 event_kind(profiler),
325 EventId::from_virtual(event_id),
333 pub fn with_profiler(&self, f: impl FnOnce(&SelfProfiler)) {
334 if let Some(profiler) = &self.profiler {
340 pub fn enabled(&self) -> bool {
341 self.profiler.is_some()
345 pub struct SelfProfiler {
347 event_filter_mask: EventFilter,
349 string_cache: RwLock<FxHashMap<&'static str, StringId>>,
351 query_event_kind: StringId,
352 generic_activity_event_kind: StringId,
353 incremental_load_result_event_kind: StringId,
354 query_blocked_event_kind: StringId,
355 query_cache_hit_event_kind: StringId,
360 output_directory: &Path,
361 crate_name: Option<&str>,
362 event_filters: &Option<Vec<String>>,
363 ) -> Result<SelfProfiler, Box<dyn Error>> {
364 fs::create_dir_all(output_directory)?;
366 let crate_name = crate_name.unwrap_or("unknown-crate");
367 let filename = format!("{}-{}.rustc_profile", crate_name, process::id());
368 let path = output_directory.join(&filename);
369 let profiler = Profiler::new(&path)?;
371 let query_event_kind = profiler.alloc_string("Query");
372 let generic_activity_event_kind = profiler.alloc_string("GenericActivity");
373 let incremental_load_result_event_kind = profiler.alloc_string("IncrementalLoadResult");
374 let query_blocked_event_kind = profiler.alloc_string("QueryBlocked");
375 let query_cache_hit_event_kind = profiler.alloc_string("QueryCacheHit");
377 let mut event_filter_mask = EventFilter::empty();
379 if let Some(ref event_filters) = *event_filters {
380 let mut unknown_events = vec![];
381 for item in event_filters {
382 if let Some(&(_, mask)) =
383 EVENT_FILTERS_BY_NAME.iter().find(|&(name, _)| name == item)
385 event_filter_mask |= mask;
387 unknown_events.push(item.clone());
391 // Warn about any unknown event names
392 if unknown_events.len() > 0 {
393 unknown_events.sort();
394 unknown_events.dedup();
397 "Unknown self-profiler events specified: {}. Available options are: {}.",
398 unknown_events.join(", "),
399 EVENT_FILTERS_BY_NAME
401 .map(|&(name, _)| name.to_string())
407 event_filter_mask = EventFilter::DEFAULT;
413 string_cache: RwLock::new(FxHashMap::default()),
415 generic_activity_event_kind,
416 incremental_load_result_event_kind,
417 query_blocked_event_kind,
418 query_cache_hit_event_kind,
422 /// Allocates a new string in the profiling data. Does not do any caching
423 /// or deduplication.
424 pub fn alloc_string<STR: SerializableString + ?Sized>(&self, s: &STR) -> StringId {
425 self.profiler.alloc_string(s)
428 /// Gets a `StringId` for the given string. This method makes sure that
429 /// any strings going through it will only be allocated once in the
431 pub fn get_or_alloc_cached_string(&self, s: &'static str) -> StringId {
432 // Only acquire a read-lock first since we assume that the string is
433 // already present in the common case.
435 let string_cache = self.string_cache.read();
437 if let Some(&id) = string_cache.get(s) {
442 let mut string_cache = self.string_cache.write();
443 // Check if the string has already been added in the small time window
444 // between dropping the read lock and acquiring the write lock.
445 *string_cache.entry(s).or_insert_with(|| self.profiler.alloc_string(s))
448 pub fn map_query_invocation_id_to_string(&self, from: QueryInvocationId, to: StringId) {
449 let from = StringId::new_virtual(from.0);
450 self.profiler.map_virtual_to_concrete_string(from, to);
453 pub fn bulk_map_query_invocation_id_to_single_string<I>(&self, from: I, to: StringId)
455 I: Iterator<Item = QueryInvocationId> + ExactSizeIterator,
457 let from = from.map(|qid| StringId::new_virtual(qid.0));
458 self.profiler.bulk_map_virtual_to_single_concrete_string(from, to);
461 pub fn query_key_recording_enabled(&self) -> bool {
462 self.event_filter_mask.contains(EventFilter::QUERY_KEYS)
465 pub fn event_id_builder(&self) -> EventIdBuilder<'_, SerializationSink> {
466 EventIdBuilder::new(&self.profiler)
471 pub struct TimingGuard<'a>(Option<measureme::TimingGuard<'a, SerializationSink>>);
473 impl<'a> TimingGuard<'a> {
476 profiler: &'a SelfProfiler,
477 event_kind: StringId,
479 ) -> TimingGuard<'a> {
480 let thread_id = thread_id_to_u32(std::thread::current().id());
481 let raw_profiler = &profiler.profiler;
483 raw_profiler.start_recording_interval_event(event_kind, event_id, thread_id);
484 TimingGuard(Some(timing_guard))
488 pub fn finish_with_query_invocation_id(self, query_invocation_id: QueryInvocationId) {
489 if let Some(guard) = self.0 {
490 let event_id = StringId::new_virtual(query_invocation_id.0);
491 let event_id = EventId::from_virtual(event_id);
492 guard.finish_with_override_event_id(event_id);
497 pub fn none() -> TimingGuard<'a> {
502 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
509 pub struct VerboseTimingGuard<'a> {
511 start: Option<Instant>,
512 _guard: TimingGuard<'a>,
515 impl<'a> VerboseTimingGuard<'a> {
516 pub fn start(event_id: &'a str, verbose: bool, _guard: TimingGuard<'a>) -> Self {
520 start: if unlikely!(verbose) { Some(Instant::now()) } else { None },
525 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
531 impl Drop for VerboseTimingGuard<'_> {
533 self.start.map(|start| print_time_passes_entry(true, self.event_id, start.elapsed()));
537 pub fn print_time_passes_entry(do_it: bool, what: &str, dur: Duration) {
542 let mem_string = match get_resident() {
544 let mb = n as f64 / 1_000_000.0;
545 format!("; rss: {}MB", mb.round() as usize)
547 None => String::new(),
549 println!("time: {}{}\t{}", duration_to_secs_str(dur), mem_string, what);
552 // Hack up our own formatting for the duration to make it easier for scripts
553 // to parse (always use the same number of decimal places and the same unit).
554 pub fn duration_to_secs_str(dur: std::time::Duration) -> String {
555 const NANOS_PER_SEC: f64 = 1_000_000_000.0;
556 let secs = dur.as_secs() as f64 + dur.subsec_nanos() as f64 / NANOS_PER_SEC;
558 format!("{:.3}", secs)
563 fn get_resident() -> Option<usize> {
565 let contents = fs::read("/proc/self/statm").ok()?;
566 let contents = String::from_utf8(contents).ok()?;
567 let s = contents.split_whitespace().nth(field)?;
568 let npages = s.parse::<usize>().ok()?;
573 fn get_resident() -> Option<usize> {
574 use std::mem::{self, MaybeUninit};
575 use winapi::shared::minwindef::DWORD;
576 use winapi::um::processthreadsapi::GetCurrentProcess;
577 use winapi::um::psapi::{GetProcessMemoryInfo, PROCESS_MEMORY_COUNTERS};
579 let mut pmc = MaybeUninit::<PROCESS_MEMORY_COUNTERS>::uninit();
581 GetProcessMemoryInfo(GetCurrentProcess(), pmc.as_mut_ptr(), mem::size_of_val(&pmc) as DWORD)
585 let pmc = unsafe { pmc.assume_init() };
586 Some(pmc.WorkingSetSize as usize)