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 use measureme::{EventId, EventIdBuilder, Profiler, SerializableString, StringId};
98 use parking_lot::RwLock;
100 bitflags::bitflags! {
101 struct EventFilter: u32 {
102 const GENERIC_ACTIVITIES = 1 << 0;
103 const QUERY_PROVIDERS = 1 << 1;
104 const QUERY_CACHE_HITS = 1 << 2;
105 const QUERY_BLOCKED = 1 << 3;
106 const INCR_CACHE_LOADS = 1 << 4;
108 const QUERY_KEYS = 1 << 5;
109 const FUNCTION_ARGS = 1 << 6;
112 const DEFAULT = Self::GENERIC_ACTIVITIES.bits |
113 Self::QUERY_PROVIDERS.bits |
114 Self::QUERY_BLOCKED.bits |
115 Self::INCR_CACHE_LOADS.bits;
117 const ARGS = Self::QUERY_KEYS.bits | Self::FUNCTION_ARGS.bits;
121 // keep this in sync with the `-Z self-profile-events` help message in librustc_session/options.rs
122 const EVENT_FILTERS_BY_NAME: &[(&str, EventFilter)] = &[
123 ("none", EventFilter::empty()),
124 ("all", EventFilter::all()),
125 ("default", EventFilter::DEFAULT),
126 ("generic-activity", EventFilter::GENERIC_ACTIVITIES),
127 ("query-provider", EventFilter::QUERY_PROVIDERS),
128 ("query-cache-hit", EventFilter::QUERY_CACHE_HITS),
129 ("query-blocked", EventFilter::QUERY_BLOCKED),
130 ("incr-cache-load", EventFilter::INCR_CACHE_LOADS),
131 ("query-keys", EventFilter::QUERY_KEYS),
132 ("function-args", EventFilter::FUNCTION_ARGS),
133 ("args", EventFilter::ARGS),
134 ("llvm", EventFilter::LLVM),
137 /// Something that uniquely identifies a query invocation.
138 pub struct QueryInvocationId(pub u32);
140 /// A reference to the SelfProfiler. It can be cloned and sent across thread
141 /// boundaries at will.
143 pub struct SelfProfilerRef {
144 // This field is `None` if self-profiling is disabled for the current
145 // compilation session.
146 profiler: Option<Arc<SelfProfiler>>,
148 // We store the filter mask directly in the reference because that doesn't
149 // cost anything and allows for filtering with checking if the profiler is
151 event_filter_mask: EventFilter,
153 // Print verbose generic activities to stdout
154 print_verbose_generic_activities: bool,
156 // Print extra verbose generic activities to stdout
157 print_extra_verbose_generic_activities: bool,
160 impl SelfProfilerRef {
162 profiler: Option<Arc<SelfProfiler>>,
163 print_verbose_generic_activities: bool,
164 print_extra_verbose_generic_activities: bool,
165 ) -> SelfProfilerRef {
166 // If there is no SelfProfiler then the filter mask is set to NONE,
167 // ensuring that nothing ever tries to actually access it.
168 let event_filter_mask =
169 profiler.as_ref().map_or(EventFilter::empty(), |p| p.event_filter_mask);
174 print_verbose_generic_activities,
175 print_extra_verbose_generic_activities,
179 // This shim makes sure that calls only get executed if the filter mask
180 // lets them pass. It also contains some trickery to make sure that
181 // code is optimized for non-profiling compilation sessions, i.e. anything
182 // past the filter check is never inlined so it doesn't clutter the fast
185 fn exec<F>(&self, event_filter: EventFilter, f: F) -> TimingGuard<'_>
187 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
190 fn cold_call<F>(profiler_ref: &SelfProfilerRef, f: F) -> TimingGuard<'_>
192 F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>,
194 let profiler = profiler_ref.profiler.as_ref().unwrap();
198 if unlikely!(self.event_filter_mask.contains(event_filter)) {
205 /// Start profiling a verbose generic activity. Profiling continues until the
206 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
207 /// a measureme event, "verbose" generic activities also print a timing entry to
208 /// stdout if the compiler is invoked with -Ztime or -Ztime-passes.
209 pub fn verbose_generic_activity<'a>(
211 event_label: &'static str,
212 ) -> VerboseTimingGuard<'a> {
214 if self.print_verbose_generic_activities { Some(event_label.to_owned()) } else { None };
216 VerboseTimingGuard::start(message, self.generic_activity(event_label))
219 /// Start profiling a extra verbose generic activity. Profiling continues until the
220 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
221 /// a measureme event, "extra verbose" generic activities also print a timing entry to
222 /// stdout if the compiler is invoked with -Ztime-passes.
223 pub fn extra_verbose_generic_activity<'a, A>(
225 event_label: &'static str,
227 ) -> VerboseTimingGuard<'a>
229 A: Borrow<str> + Into<String>,
231 let message = if self.print_extra_verbose_generic_activities {
232 Some(format!("{}({})", event_label, event_arg.borrow()))
237 VerboseTimingGuard::start(message, self.generic_activity_with_arg(event_label, event_arg))
240 /// Start profiling a generic activity. Profiling continues until the
241 /// TimingGuard returned from this call is dropped.
243 pub fn generic_activity(&self, event_label: &'static str) -> TimingGuard<'_> {
244 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
245 let event_label = profiler.get_or_alloc_cached_string(event_label);
246 let event_id = EventId::from_label(event_label);
247 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
251 /// Start profiling a generic activity. Profiling continues until the
252 /// TimingGuard returned from this call is dropped.
254 pub fn generic_activity_with_arg<A>(
256 event_label: &'static str,
260 A: Borrow<str> + Into<String>,
262 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
263 let builder = EventIdBuilder::new(&profiler.profiler);
264 let event_label = profiler.get_or_alloc_cached_string(event_label);
265 let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) {
266 let event_arg = profiler.get_or_alloc_cached_string(event_arg);
267 builder.from_label_and_arg(event_label, event_arg)
269 builder.from_label(event_label)
271 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
276 pub fn generic_activity_with_args(
278 event_label: &'static str,
279 event_args: &[String],
280 ) -> TimingGuard<'_> {
281 self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| {
282 let builder = EventIdBuilder::new(&profiler.profiler);
283 let event_label = profiler.get_or_alloc_cached_string(event_label);
284 let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) {
285 let event_args: Vec<_> = event_args
287 .map(|s| profiler.get_or_alloc_cached_string(&s[..]))
289 builder.from_label_and_args(event_label, &event_args)
291 builder.from_label(event_label)
293 TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id)
297 /// Start profiling a query provider. Profiling continues until the
298 /// TimingGuard returned from this call is dropped.
300 pub fn query_provider(&self) -> TimingGuard<'_> {
301 self.exec(EventFilter::QUERY_PROVIDERS, |profiler| {
302 TimingGuard::start(profiler, profiler.query_event_kind, EventId::INVALID)
306 /// Record a query in-memory cache hit.
308 pub fn query_cache_hit(&self, query_invocation_id: QueryInvocationId) {
309 self.instant_query_event(
310 |profiler| profiler.query_cache_hit_event_kind,
312 EventFilter::QUERY_CACHE_HITS,
316 /// Start profiling a query being blocked on a concurrent execution.
317 /// Profiling continues until the TimingGuard returned from this call is
320 pub fn query_blocked(&self) -> TimingGuard<'_> {
321 self.exec(EventFilter::QUERY_BLOCKED, |profiler| {
322 TimingGuard::start(profiler, profiler.query_blocked_event_kind, EventId::INVALID)
326 /// Start profiling how long it takes to load a query result from the
327 /// incremental compilation on-disk cache. Profiling continues until the
328 /// TimingGuard returned from this call is dropped.
330 pub fn incr_cache_loading(&self) -> TimingGuard<'_> {
331 self.exec(EventFilter::INCR_CACHE_LOADS, |profiler| {
334 profiler.incremental_load_result_event_kind,
341 fn instant_query_event(
343 event_kind: fn(&SelfProfiler) -> StringId,
344 query_invocation_id: QueryInvocationId,
345 event_filter: EventFilter,
347 drop(self.exec(event_filter, |profiler| {
348 let event_id = StringId::new_virtual(query_invocation_id.0);
349 let thread_id = std::thread::current().id().as_u64().get() as u32;
351 profiler.profiler.record_instant_event(
352 event_kind(profiler),
353 EventId::from_virtual(event_id),
361 pub fn with_profiler(&self, f: impl FnOnce(&SelfProfiler)) {
362 if let Some(profiler) = &self.profiler {
368 pub fn enabled(&self) -> bool {
369 self.profiler.is_some()
373 pub fn llvm_recording_enabled(&self) -> bool {
374 self.event_filter_mask.contains(EventFilter::LLVM)
377 pub fn get_self_profiler(&self) -> Option<Arc<SelfProfiler>> {
378 self.profiler.clone()
382 pub struct SelfProfiler {
384 event_filter_mask: EventFilter,
386 string_cache: RwLock<FxHashMap<String, StringId>>,
388 query_event_kind: StringId,
389 generic_activity_event_kind: StringId,
390 incremental_load_result_event_kind: StringId,
391 query_blocked_event_kind: StringId,
392 query_cache_hit_event_kind: StringId,
397 output_directory: &Path,
398 crate_name: Option<&str>,
399 event_filters: &Option<Vec<String>>,
400 ) -> Result<SelfProfiler, Box<dyn Error + Send + Sync>> {
401 fs::create_dir_all(output_directory)?;
403 let crate_name = crate_name.unwrap_or("unknown-crate");
404 let filename = format!("{}-{}.rustc_profile", crate_name, process::id());
405 let path = output_directory.join(&filename);
406 let profiler = Profiler::new(&path)?;
408 let query_event_kind = profiler.alloc_string("Query");
409 let generic_activity_event_kind = profiler.alloc_string("GenericActivity");
410 let incremental_load_result_event_kind = profiler.alloc_string("IncrementalLoadResult");
411 let query_blocked_event_kind = profiler.alloc_string("QueryBlocked");
412 let query_cache_hit_event_kind = profiler.alloc_string("QueryCacheHit");
414 let mut event_filter_mask = EventFilter::empty();
416 if let Some(ref event_filters) = *event_filters {
417 let mut unknown_events = vec![];
418 for item in event_filters {
419 if let Some(&(_, mask)) =
420 EVENT_FILTERS_BY_NAME.iter().find(|&(name, _)| name == item)
422 event_filter_mask |= mask;
424 unknown_events.push(item.clone());
428 // Warn about any unknown event names
429 if !unknown_events.is_empty() {
430 unknown_events.sort();
431 unknown_events.dedup();
434 "Unknown self-profiler events specified: {}. Available options are: {}.",
435 unknown_events.join(", "),
436 EVENT_FILTERS_BY_NAME
438 .map(|&(name, _)| name.to_string())
444 event_filter_mask = EventFilter::DEFAULT;
450 string_cache: RwLock::new(FxHashMap::default()),
452 generic_activity_event_kind,
453 incremental_load_result_event_kind,
454 query_blocked_event_kind,
455 query_cache_hit_event_kind,
459 /// Allocates a new string in the profiling data. Does not do any caching
460 /// or deduplication.
461 pub fn alloc_string<STR: SerializableString + ?Sized>(&self, s: &STR) -> StringId {
462 self.profiler.alloc_string(s)
465 /// Gets a `StringId` for the given string. This method makes sure that
466 /// any strings going through it will only be allocated once in the
468 pub fn get_or_alloc_cached_string<A>(&self, s: A) -> StringId
470 A: Borrow<str> + Into<String>,
472 // Only acquire a read-lock first since we assume that the string is
473 // already present in the common case.
475 let string_cache = self.string_cache.read();
477 if let Some(&id) = string_cache.get(s.borrow()) {
482 let mut string_cache = self.string_cache.write();
483 // Check if the string has already been added in the small time window
484 // between dropping the read lock and acquiring the write lock.
485 match string_cache.entry(s.into()) {
486 Entry::Occupied(e) => *e.get(),
487 Entry::Vacant(e) => {
488 let string_id = self.profiler.alloc_string(&e.key()[..]);
494 pub fn map_query_invocation_id_to_string(&self, from: QueryInvocationId, to: StringId) {
495 let from = StringId::new_virtual(from.0);
496 self.profiler.map_virtual_to_concrete_string(from, to);
499 pub fn bulk_map_query_invocation_id_to_single_string<I>(&self, from: I, to: StringId)
501 I: Iterator<Item = QueryInvocationId> + ExactSizeIterator,
503 let from = from.map(|qid| StringId::new_virtual(qid.0));
504 self.profiler.bulk_map_virtual_to_single_concrete_string(from, to);
507 pub fn query_key_recording_enabled(&self) -> bool {
508 self.event_filter_mask.contains(EventFilter::QUERY_KEYS)
511 pub fn event_id_builder(&self) -> EventIdBuilder<'_> {
512 EventIdBuilder::new(&self.profiler)
517 pub struct TimingGuard<'a>(Option<measureme::TimingGuard<'a>>);
519 impl<'a> TimingGuard<'a> {
522 profiler: &'a SelfProfiler,
523 event_kind: StringId,
525 ) -> TimingGuard<'a> {
526 let thread_id = std::thread::current().id().as_u64().get() as u32;
527 let raw_profiler = &profiler.profiler;
529 raw_profiler.start_recording_interval_event(event_kind, event_id, thread_id);
530 TimingGuard(Some(timing_guard))
534 pub fn finish_with_query_invocation_id(self, query_invocation_id: QueryInvocationId) {
535 if let Some(guard) = self.0 {
537 let event_id = StringId::new_virtual(query_invocation_id.0);
538 let event_id = EventId::from_virtual(event_id);
539 guard.finish_with_override_event_id(event_id);
545 pub fn none() -> TimingGuard<'a> {
550 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
557 pub struct VerboseTimingGuard<'a> {
558 start_and_message: Option<(Instant, Option<usize>, String)>,
559 _guard: TimingGuard<'a>,
562 impl<'a> VerboseTimingGuard<'a> {
563 pub fn start(message: Option<String>, _guard: TimingGuard<'a>) -> Self {
566 start_and_message: message.map(|msg| (Instant::now(), get_resident_set_size(), msg)),
571 pub fn run<R>(self, f: impl FnOnce() -> R) -> R {
577 impl Drop for VerboseTimingGuard<'_> {
579 if let Some((start_time, start_rss, ref message)) = self.start_and_message {
580 let end_rss = get_resident_set_size();
581 print_time_passes_entry(&message[..], start_time.elapsed(), start_rss, end_rss);
586 pub fn print_time_passes_entry(
589 start_rss: Option<usize>,
590 end_rss: Option<usize>,
592 let rss_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as usize;
593 let rss_change_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as i128;
595 let mem_string = match (start_rss, end_rss) {
596 (Some(start_rss), Some(end_rss)) => {
597 let change_rss = end_rss as i128 - start_rss as i128;
600 "; rss: {:>4}MB -> {:>4}MB ({:>+5}MB)",
601 rss_to_mb(start_rss),
603 rss_change_to_mb(change_rss),
606 (Some(start_rss), None) => format!("; rss start: {:>4}MB", rss_to_mb(start_rss)),
607 (None, Some(end_rss)) => format!("; rss end: {:>4}MB", rss_to_mb(end_rss)),
608 (None, None) => String::new(),
611 eprintln!("time: {:>7}{}\t{}", duration_to_secs_str(dur), mem_string, what);
614 // Hack up our own formatting for the duration to make it easier for scripts
615 // to parse (always use the same number of decimal places and the same unit).
616 pub fn duration_to_secs_str(dur: std::time::Duration) -> String {
617 format!("{:.3}", dur.as_secs_f64())
623 pub fn get_resident_set_size() -> Option<usize> {
624 use std::mem::{self, MaybeUninit};
625 use winapi::shared::minwindef::DWORD;
626 use winapi::um::processthreadsapi::GetCurrentProcess;
627 use winapi::um::psapi::{GetProcessMemoryInfo, PROCESS_MEMORY_COUNTERS};
629 let mut pmc = MaybeUninit::<PROCESS_MEMORY_COUNTERS>::uninit();
631 GetProcessMemoryInfo(GetCurrentProcess(), pmc.as_mut_ptr(), mem::size_of_val(&pmc) as DWORD)
635 let pmc = unsafe { pmc.assume_init() };
636 Some(pmc.WorkingSetSize as usize)
640 } else if #[cfg(unix)] {
641 pub fn get_resident_set_size() -> Option<usize> {
643 let contents = fs::read("/proc/self/statm").ok()?;
644 let contents = String::from_utf8(contents).ok()?;
645 let s = contents.split_whitespace().nth(field)?;
646 let npages = s.parse::<usize>().ok()?;
650 pub fn get_resident_set_size() -> Option<usize> {