]>
Commit | Line | Data |
---|---|---|
dfeec247 XL |
1 | //! # Rust Compiler Self-Profiling |
2 | //! | |
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. | |
8 | //! | |
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. | |
14 | //! | |
15 | //! | |
16 | //! ## Event Overview | |
17 | //! | |
18 | //! Events have a few properties: | |
19 | //! | |
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. | |
28 | //! | |
29 | //! | |
30 | //! ## Event Filtering | |
31 | //! | |
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. | |
37 | //! | |
38 | //! | |
39 | //! ## `event_id` Assignment | |
40 | //! | |
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: | |
45 | //! | |
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. | |
53 | //! | |
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. | |
61 | //! | |
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()`. | |
66 | //! | |
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. | |
72 | //! | |
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`. | |
81 | //! | |
82 | //! [mm]: https://github.com/rust-lang/measureme/ | |
83 | ||
74b04a01 | 84 | use crate::cold_path; |
dfeec247 XL |
85 | use crate::fx::FxHashMap; |
86 | ||
74b04a01 XL |
87 | use std::borrow::Borrow; |
88 | use std::collections::hash_map::Entry; | |
48663c56 | 89 | use std::error::Error; |
dc9dc135 | 90 | use std::fs; |
dc9dc135 | 91 | use std::path::Path; |
532ac7d7 | 92 | use std::process; |
e74abb32 | 93 | use std::sync::Arc; |
dfeec247 | 94 | use std::time::{Duration, Instant}; |
b7449926 | 95 | |
136023e0 XL |
96 | pub use measureme::EventId; |
97 | use measureme::{EventIdBuilder, Profiler, SerializableString, StringId}; | |
dfeec247 | 98 | use parking_lot::RwLock; |
04454e1e | 99 | use smallvec::SmallVec; |
48663c56 | 100 | |
60c5eb7d | 101 | bitflags::bitflags! { |
48663c56 | 102 | struct EventFilter: u32 { |
136023e0 XL |
103 | const GENERIC_ACTIVITIES = 1 << 0; |
104 | const QUERY_PROVIDERS = 1 << 1; | |
105 | const QUERY_CACHE_HITS = 1 << 2; | |
106 | const QUERY_BLOCKED = 1 << 3; | |
107 | const INCR_CACHE_LOADS = 1 << 4; | |
48663c56 | 108 | |
136023e0 XL |
109 | const QUERY_KEYS = 1 << 5; |
110 | const FUNCTION_ARGS = 1 << 6; | |
111 | const LLVM = 1 << 7; | |
112 | const INCR_RESULT_HASHING = 1 << 8; | |
3c0e092e | 113 | const ARTIFACT_SIZES = 1 << 9; |
dfeec247 | 114 | |
48663c56 XL |
115 | const DEFAULT = Self::GENERIC_ACTIVITIES.bits | |
116 | Self::QUERY_PROVIDERS.bits | | |
117 | Self::QUERY_BLOCKED.bits | | |
136023e0 | 118 | Self::INCR_CACHE_LOADS.bits | |
3c0e092e XL |
119 | Self::INCR_RESULT_HASHING.bits | |
120 | Self::ARTIFACT_SIZES.bits; | |
74b04a01 XL |
121 | |
122 | const ARGS = Self::QUERY_KEYS.bits | Self::FUNCTION_ARGS.bits; | |
9fa01778 XL |
123 | } |
124 | } | |
b7449926 | 125 | |
136023e0 | 126 | // keep this in sync with the `-Z self-profile-events` help message in rustc_session/options.rs |
48663c56 | 127 | const EVENT_FILTERS_BY_NAME: &[(&str, EventFilter)] = &[ |
dfeec247 XL |
128 | ("none", EventFilter::empty()), |
129 | ("all", EventFilter::all()), | |
130 | ("default", EventFilter::DEFAULT), | |
48663c56 XL |
131 | ("generic-activity", EventFilter::GENERIC_ACTIVITIES), |
132 | ("query-provider", EventFilter::QUERY_PROVIDERS), | |
133 | ("query-cache-hit", EventFilter::QUERY_CACHE_HITS), | |
dfeec247 | 134 | ("query-blocked", EventFilter::QUERY_BLOCKED), |
48663c56 | 135 | ("incr-cache-load", EventFilter::INCR_CACHE_LOADS), |
dfeec247 | 136 | ("query-keys", EventFilter::QUERY_KEYS), |
74b04a01 XL |
137 | ("function-args", EventFilter::FUNCTION_ARGS), |
138 | ("args", EventFilter::ARGS), | |
139 | ("llvm", EventFilter::LLVM), | |
136023e0 | 140 | ("incr-result-hashing", EventFilter::INCR_RESULT_HASHING), |
3c0e092e | 141 | ("artifact-sizes", EventFilter::ARTIFACT_SIZES), |
48663c56 XL |
142 | ]; |
143 | ||
dfeec247 XL |
144 | /// Something that uniquely identifies a query invocation. |
145 | pub struct QueryInvocationId(pub u32); | |
e74abb32 XL |
146 | |
147 | /// A reference to the SelfProfiler. It can be cloned and sent across thread | |
148 | /// boundaries at will. | |
149 | #[derive(Clone)] | |
150 | pub struct SelfProfilerRef { | |
151 | // This field is `None` if self-profiling is disabled for the current | |
152 | // compilation session. | |
153 | profiler: Option<Arc<SelfProfiler>>, | |
154 | ||
155 | // We store the filter mask directly in the reference because that doesn't | |
156 | // cost anything and allows for filtering with checking if the profiler is | |
157 | // actually enabled. | |
158 | event_filter_mask: EventFilter, | |
dfeec247 | 159 | |
2b03887a | 160 | // Print verbose generic activities to stderr? |
dfeec247 | 161 | print_verbose_generic_activities: bool, |
e74abb32 XL |
162 | } |
163 | ||
164 | impl SelfProfilerRef { | |
dfeec247 XL |
165 | pub fn new( |
166 | profiler: Option<Arc<SelfProfiler>>, | |
167 | print_verbose_generic_activities: bool, | |
dfeec247 | 168 | ) -> SelfProfilerRef { |
e74abb32 XL |
169 | // If there is no SelfProfiler then the filter mask is set to NONE, |
170 | // ensuring that nothing ever tries to actually access it. | |
dfeec247 | 171 | let event_filter_mask = |
5869c6ff | 172 | profiler.as_ref().map_or(EventFilter::empty(), |p| p.event_filter_mask); |
e74abb32 | 173 | |
2b03887a | 174 | SelfProfilerRef { profiler, event_filter_mask, print_verbose_generic_activities } |
e74abb32 XL |
175 | } |
176 | ||
04454e1e FG |
177 | /// This shim makes sure that calls only get executed if the filter mask |
178 | /// lets them pass. It also contains some trickery to make sure that | |
179 | /// code is optimized for non-profiling compilation sessions, i.e. anything | |
180 | /// past the filter check is never inlined so it doesn't clutter the fast | |
181 | /// path. | |
e74abb32 XL |
182 | #[inline(always)] |
183 | fn exec<F>(&self, event_filter: EventFilter, f: F) -> TimingGuard<'_> | |
dfeec247 XL |
184 | where |
185 | F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>, | |
e74abb32 XL |
186 | { |
187 | #[inline(never)] | |
923072b8 | 188 | #[cold] |
e74abb32 | 189 | fn cold_call<F>(profiler_ref: &SelfProfilerRef, f: F) -> TimingGuard<'_> |
dfeec247 XL |
190 | where |
191 | F: for<'a> FnOnce(&'a SelfProfiler) -> TimingGuard<'a>, | |
e74abb32 XL |
192 | { |
193 | let profiler = profiler_ref.profiler.as_ref().unwrap(); | |
487cf647 | 194 | f(profiler) |
e74abb32 XL |
195 | } |
196 | ||
923072b8 | 197 | if self.event_filter_mask.contains(event_filter) { |
e74abb32 XL |
198 | cold_call(self, f) |
199 | } else { | |
200 | TimingGuard::none() | |
201 | } | |
202 | } | |
203 | ||
dfeec247 XL |
204 | /// Start profiling a verbose generic activity. Profiling continues until the |
205 | /// VerboseTimingGuard returned from this call is dropped. In addition to recording | |
206 | /// a measureme event, "verbose" generic activities also print a timing entry to | |
2b03887a | 207 | /// stderr if the compiler is invoked with -Ztime-passes. |
9c376795 | 208 | pub fn verbose_generic_activity(&self, event_label: &'static str) -> VerboseTimingGuard<'_> { |
74b04a01 XL |
209 | let message = |
210 | if self.print_verbose_generic_activities { Some(event_label.to_owned()) } else { None }; | |
211 | ||
212 | VerboseTimingGuard::start(message, self.generic_activity(event_label)) | |
dfeec247 XL |
213 | } |
214 | ||
2b03887a | 215 | /// Like `verbose_generic_activity`, but with an extra arg. |
9c376795 FG |
216 | pub fn verbose_generic_activity_with_arg<A>( |
217 | &self, | |
74b04a01 XL |
218 | event_label: &'static str, |
219 | event_arg: A, | |
9c376795 | 220 | ) -> VerboseTimingGuard<'_> |
74b04a01 XL |
221 | where |
222 | A: Borrow<str> + Into<String>, | |
223 | { | |
2b03887a | 224 | let message = if self.print_verbose_generic_activities { |
74b04a01 XL |
225 | Some(format!("{}({})", event_label, event_arg.borrow())) |
226 | } else { | |
227 | None | |
228 | }; | |
229 | ||
230 | VerboseTimingGuard::start(message, self.generic_activity_with_arg(event_label, event_arg)) | |
231 | } | |
232 | ||
233 | /// Start profiling a generic activity. Profiling continues until the | |
234 | /// TimingGuard returned from this call is dropped. | |
235 | #[inline(always)] | |
236 | pub fn generic_activity(&self, event_label: &'static str) -> TimingGuard<'_> { | |
237 | self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| { | |
238 | let event_label = profiler.get_or_alloc_cached_string(event_label); | |
239 | let event_id = EventId::from_label(event_label); | |
240 | TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id) | |
241 | }) | |
dfeec247 XL |
242 | } |
243 | ||
136023e0 XL |
244 | /// Start profiling with some event filter for a given event. Profiling continues until the |
245 | /// TimingGuard returned from this call is dropped. | |
246 | #[inline(always)] | |
247 | pub fn generic_activity_with_event_id(&self, event_id: EventId) -> TimingGuard<'_> { | |
248 | self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| { | |
249 | TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id) | |
250 | }) | |
251 | } | |
252 | ||
e74abb32 XL |
253 | /// Start profiling a generic activity. Profiling continues until the |
254 | /// TimingGuard returned from this call is dropped. | |
255 | #[inline(always)] | |
74b04a01 XL |
256 | pub fn generic_activity_with_arg<A>( |
257 | &self, | |
258 | event_label: &'static str, | |
259 | event_arg: A, | |
260 | ) -> TimingGuard<'_> | |
261 | where | |
262 | A: Borrow<str> + Into<String>, | |
263 | { | |
e74abb32 | 264 | self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| { |
74b04a01 XL |
265 | let builder = EventIdBuilder::new(&profiler.profiler); |
266 | let event_label = profiler.get_or_alloc_cached_string(event_label); | |
267 | let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) { | |
268 | let event_arg = profiler.get_or_alloc_cached_string(event_arg); | |
269 | builder.from_label_and_arg(event_label, event_arg) | |
270 | } else { | |
271 | builder.from_label(event_label) | |
272 | }; | |
dfeec247 | 273 | TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id) |
e74abb32 XL |
274 | }) |
275 | } | |
276 | ||
04454e1e FG |
277 | /// Start profiling a generic activity, allowing costly arguments to be recorded. Profiling |
278 | /// continues until the `TimingGuard` returned from this call is dropped. | |
279 | /// | |
280 | /// If the arguments to a generic activity are cheap to create, use `generic_activity_with_arg` | |
281 | /// or `generic_activity_with_args` for their simpler API. However, if they are costly or | |
282 | /// require allocation in sufficiently hot contexts, then this allows for a closure to be called | |
283 | /// only when arguments were asked to be recorded via `-Z self-profile-events=args`. | |
284 | /// | |
285 | /// In this case, the closure will be passed a `&mut EventArgRecorder`, to help with recording | |
286 | /// one or many arguments within the generic activity being profiled, by calling its | |
287 | /// `record_arg` method for example. | |
288 | /// | |
289 | /// This `EventArgRecorder` may implement more specific traits from other rustc crates, e.g. for | |
290 | /// richer handling of rustc-specific argument types, while keeping this single entry-point API | |
291 | /// for recording arguments. | |
292 | /// | |
293 | /// Note: recording at least one argument is *required* for the self-profiler to create the | |
294 | /// `TimingGuard`. A panic will be triggered if that doesn't happen. This function exists | |
295 | /// explicitly to record arguments, so it fails loudly when there are none to record. | |
296 | /// | |
297 | #[inline(always)] | |
298 | pub fn generic_activity_with_arg_recorder<F>( | |
299 | &self, | |
300 | event_label: &'static str, | |
301 | mut f: F, | |
302 | ) -> TimingGuard<'_> | |
303 | where | |
304 | F: FnMut(&mut EventArgRecorder<'_>), | |
305 | { | |
306 | // Ensure this event will only be recorded when self-profiling is turned on. | |
307 | self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| { | |
308 | let builder = EventIdBuilder::new(&profiler.profiler); | |
309 | let event_label = profiler.get_or_alloc_cached_string(event_label); | |
310 | ||
311 | // Ensure the closure to create event arguments will only be called when argument | |
312 | // recording is turned on. | |
313 | let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) { | |
314 | // Set up the builder and call the user-provided closure to record potentially | |
315 | // costly event arguments. | |
316 | let mut recorder = EventArgRecorder { profiler, args: SmallVec::new() }; | |
317 | f(&mut recorder); | |
318 | ||
319 | // It is expected that the closure will record at least one argument. If that | |
320 | // doesn't happen, it's a bug: we've been explicitly called in order to record | |
321 | // arguments, so we fail loudly when there are none to record. | |
322 | if recorder.args.is_empty() { | |
323 | panic!( | |
324 | "The closure passed to `generic_activity_with_arg_recorder` needs to \ | |
325 | record at least one argument" | |
326 | ); | |
327 | } | |
328 | ||
329 | builder.from_label_and_args(event_label, &recorder.args) | |
330 | } else { | |
331 | builder.from_label(event_label) | |
332 | }; | |
333 | TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id) | |
334 | }) | |
335 | } | |
336 | ||
3c0e092e XL |
337 | /// Record the size of an artifact that the compiler produces |
338 | /// | |
339 | /// `artifact_kind` is the class of artifact (e.g., query_cache, object_file, etc.) | |
340 | /// `artifact_name` is an identifier to the specific artifact being stored (usually a filename) | |
341 | #[inline(always)] | |
342 | pub fn artifact_size<A>(&self, artifact_kind: &str, artifact_name: A, size: u64) | |
343 | where | |
344 | A: Borrow<str> + Into<String>, | |
345 | { | |
346 | drop(self.exec(EventFilter::ARTIFACT_SIZES, |profiler| { | |
347 | let builder = EventIdBuilder::new(&profiler.profiler); | |
348 | let event_label = profiler.get_or_alloc_cached_string(artifact_kind); | |
349 | let event_arg = profiler.get_or_alloc_cached_string(artifact_name); | |
350 | let event_id = builder.from_label_and_arg(event_label, event_arg); | |
351 | let thread_id = get_thread_id(); | |
352 | ||
353 | profiler.profiler.record_integer_event( | |
354 | profiler.artifact_size_event_kind, | |
355 | event_id, | |
356 | thread_id, | |
357 | size, | |
358 | ); | |
359 | ||
360 | TimingGuard::none() | |
361 | })) | |
362 | } | |
363 | ||
fc512014 XL |
364 | #[inline(always)] |
365 | pub fn generic_activity_with_args( | |
366 | &self, | |
367 | event_label: &'static str, | |
368 | event_args: &[String], | |
369 | ) -> TimingGuard<'_> { | |
370 | self.exec(EventFilter::GENERIC_ACTIVITIES, |profiler| { | |
371 | let builder = EventIdBuilder::new(&profiler.profiler); | |
372 | let event_label = profiler.get_or_alloc_cached_string(event_label); | |
373 | let event_id = if profiler.event_filter_mask.contains(EventFilter::FUNCTION_ARGS) { | |
374 | let event_args: Vec<_> = event_args | |
375 | .iter() | |
376 | .map(|s| profiler.get_or_alloc_cached_string(&s[..])) | |
377 | .collect(); | |
378 | builder.from_label_and_args(event_label, &event_args) | |
379 | } else { | |
380 | builder.from_label(event_label) | |
381 | }; | |
382 | TimingGuard::start(profiler, profiler.generic_activity_event_kind, event_id) | |
383 | }) | |
384 | } | |
385 | ||
e74abb32 XL |
386 | /// Start profiling a query provider. Profiling continues until the |
387 | /// TimingGuard returned from this call is dropped. | |
388 | #[inline(always)] | |
dfeec247 | 389 | pub fn query_provider(&self) -> TimingGuard<'_> { |
e74abb32 | 390 | self.exec(EventFilter::QUERY_PROVIDERS, |profiler| { |
dfeec247 | 391 | TimingGuard::start(profiler, profiler.query_event_kind, EventId::INVALID) |
e74abb32 XL |
392 | }) |
393 | } | |
394 | ||
395 | /// Record a query in-memory cache hit. | |
396 | #[inline(always)] | |
dfeec247 | 397 | pub fn query_cache_hit(&self, query_invocation_id: QueryInvocationId) { |
60c5eb7d | 398 | self.instant_query_event( |
e74abb32 | 399 | |profiler| profiler.query_cache_hit_event_kind, |
dfeec247 | 400 | query_invocation_id, |
e74abb32 | 401 | EventFilter::QUERY_CACHE_HITS, |
e74abb32 XL |
402 | ); |
403 | } | |
404 | ||
405 | /// Start profiling a query being blocked on a concurrent execution. | |
406 | /// Profiling continues until the TimingGuard returned from this call is | |
407 | /// dropped. | |
408 | #[inline(always)] | |
dfeec247 | 409 | pub fn query_blocked(&self) -> TimingGuard<'_> { |
e74abb32 | 410 | self.exec(EventFilter::QUERY_BLOCKED, |profiler| { |
dfeec247 | 411 | TimingGuard::start(profiler, profiler.query_blocked_event_kind, EventId::INVALID) |
e74abb32 XL |
412 | }) |
413 | } | |
414 | ||
415 | /// Start profiling how long it takes to load a query result from the | |
416 | /// incremental compilation on-disk cache. Profiling continues until the | |
417 | /// TimingGuard returned from this call is dropped. | |
418 | #[inline(always)] | |
dfeec247 | 419 | pub fn incr_cache_loading(&self) -> TimingGuard<'_> { |
e74abb32 | 420 | self.exec(EventFilter::INCR_CACHE_LOADS, |profiler| { |
e74abb32 XL |
421 | TimingGuard::start( |
422 | profiler, | |
423 | profiler.incremental_load_result_event_kind, | |
dfeec247 | 424 | EventId::INVALID, |
e74abb32 XL |
425 | ) |
426 | }) | |
427 | } | |
428 | ||
136023e0 XL |
429 | /// Start profiling how long it takes to hash query results for incremental compilation. |
430 | /// Profiling continues until the TimingGuard returned from this call is dropped. | |
431 | #[inline(always)] | |
432 | pub fn incr_result_hashing(&self) -> TimingGuard<'_> { | |
433 | self.exec(EventFilter::INCR_RESULT_HASHING, |profiler| { | |
434 | TimingGuard::start( | |
435 | profiler, | |
436 | profiler.incremental_result_hashing_event_kind, | |
437 | EventId::INVALID, | |
438 | ) | |
439 | }) | |
440 | } | |
441 | ||
e74abb32 | 442 | #[inline(always)] |
60c5eb7d | 443 | fn instant_query_event( |
e74abb32 XL |
444 | &self, |
445 | event_kind: fn(&SelfProfiler) -> StringId, | |
dfeec247 | 446 | query_invocation_id: QueryInvocationId, |
e74abb32 | 447 | event_filter: EventFilter, |
e74abb32 XL |
448 | ) { |
449 | drop(self.exec(event_filter, |profiler| { | |
dfeec247 | 450 | let event_id = StringId::new_virtual(query_invocation_id.0); |
3c0e092e | 451 | let thread_id = get_thread_id(); |
e74abb32 | 452 | |
60c5eb7d | 453 | profiler.profiler.record_instant_event( |
e74abb32 | 454 | event_kind(profiler), |
dfeec247 | 455 | EventId::from_virtual(event_id), |
e74abb32 | 456 | thread_id, |
e74abb32 XL |
457 | ); |
458 | ||
459 | TimingGuard::none() | |
460 | })); | |
461 | } | |
60c5eb7d | 462 | |
dfeec247 | 463 | pub fn with_profiler(&self, f: impl FnOnce(&SelfProfiler)) { |
60c5eb7d | 464 | if let Some(profiler) = &self.profiler { |
487cf647 | 465 | f(profiler) |
60c5eb7d XL |
466 | } |
467 | } | |
dfeec247 | 468 | |
136023e0 XL |
469 | /// Gets a `StringId` for the given string. This method makes sure that |
470 | /// any strings going through it will only be allocated once in the | |
471 | /// profiling data. | |
472 | /// Returns `None` if the self-profiling is not enabled. | |
473 | pub fn get_or_alloc_cached_string(&self, s: &str) -> Option<StringId> { | |
474 | self.profiler.as_ref().map(|p| p.get_or_alloc_cached_string(s)) | |
475 | } | |
476 | ||
dfeec247 XL |
477 | #[inline] |
478 | pub fn enabled(&self) -> bool { | |
479 | self.profiler.is_some() | |
480 | } | |
74b04a01 XL |
481 | |
482 | #[inline] | |
483 | pub fn llvm_recording_enabled(&self) -> bool { | |
484 | self.event_filter_mask.contains(EventFilter::LLVM) | |
485 | } | |
486 | #[inline] | |
487 | pub fn get_self_profiler(&self) -> Option<Arc<SelfProfiler>> { | |
488 | self.profiler.clone() | |
489 | } | |
e74abb32 XL |
490 | } |
491 | ||
04454e1e FG |
492 | /// A helper for recording costly arguments to self-profiling events. Used with |
493 | /// `SelfProfilerRef::generic_activity_with_arg_recorder`. | |
494 | pub struct EventArgRecorder<'p> { | |
495 | /// The `SelfProfiler` used to intern the event arguments that users will ask to record. | |
496 | profiler: &'p SelfProfiler, | |
497 | ||
498 | /// The interned event arguments to be recorded in the generic activity event. | |
499 | /// | |
500 | /// The most common case, when actually recording event arguments, is to have one argument. Then | |
501 | /// followed by recording two, in a couple places. | |
502 | args: SmallVec<[StringId; 2]>, | |
503 | } | |
504 | ||
505 | impl EventArgRecorder<'_> { | |
506 | /// Records a single argument within the current generic activity being profiled. | |
507 | /// | |
508 | /// Note: when self-profiling with costly event arguments, at least one argument | |
509 | /// needs to be recorded. A panic will be triggered if that doesn't happen. | |
510 | pub fn record_arg<A>(&mut self, event_arg: A) | |
511 | where | |
512 | A: Borrow<str> + Into<String>, | |
513 | { | |
514 | let event_arg = self.profiler.get_or_alloc_cached_string(event_arg); | |
515 | self.args.push(event_arg); | |
516 | } | |
517 | } | |
518 | ||
532ac7d7 | 519 | pub struct SelfProfiler { |
48663c56 XL |
520 | profiler: Profiler, |
521 | event_filter_mask: EventFilter, | |
dfeec247 | 522 | |
74b04a01 | 523 | string_cache: RwLock<FxHashMap<String, StringId>>, |
dfeec247 | 524 | |
48663c56 XL |
525 | query_event_kind: StringId, |
526 | generic_activity_event_kind: StringId, | |
527 | incremental_load_result_event_kind: StringId, | |
136023e0 | 528 | incremental_result_hashing_event_kind: StringId, |
48663c56 XL |
529 | query_blocked_event_kind: StringId, |
530 | query_cache_hit_event_kind: StringId, | |
3c0e092e | 531 | artifact_size_event_kind: StringId, |
b7449926 XL |
532 | } |
533 | ||
534 | impl SelfProfiler { | |
dc9dc135 XL |
535 | pub fn new( |
536 | output_directory: &Path, | |
537 | crate_name: Option<&str>, | |
923072b8 FG |
538 | event_filters: Option<&[String]>, |
539 | counter_name: &str, | |
29967ef6 | 540 | ) -> Result<SelfProfiler, Box<dyn Error + Send + Sync>> { |
dc9dc135 XL |
541 | fs::create_dir_all(output_directory)?; |
542 | ||
543 | let crate_name = crate_name.unwrap_or("unknown-crate"); | |
923072b8 FG |
544 | // HACK(eddyb) we need to pad the PID, strange as it may seem, as its |
545 | // length can behave as a source of entropy for heap addresses, when | |
546 | // ASLR is disabled and the heap is otherwise determinic. | |
547 | let pid: u32 = process::id(); | |
9c376795 | 548 | let filename = format!("{crate_name}-{pid:07}.rustc_profile"); |
dc9dc135 | 549 | let path = output_directory.join(&filename); |
923072b8 FG |
550 | let profiler = |
551 | Profiler::with_counter(&path, measureme::counters::Counter::by_name(counter_name)?)?; | |
48663c56 XL |
552 | |
553 | let query_event_kind = profiler.alloc_string("Query"); | |
554 | let generic_activity_event_kind = profiler.alloc_string("GenericActivity"); | |
555 | let incremental_load_result_event_kind = profiler.alloc_string("IncrementalLoadResult"); | |
136023e0 XL |
556 | let incremental_result_hashing_event_kind = |
557 | profiler.alloc_string("IncrementalResultHashing"); | |
48663c56 XL |
558 | let query_blocked_event_kind = profiler.alloc_string("QueryBlocked"); |
559 | let query_cache_hit_event_kind = profiler.alloc_string("QueryCacheHit"); | |
3c0e092e | 560 | let artifact_size_event_kind = profiler.alloc_string("ArtifactSize"); |
48663c56 XL |
561 | |
562 | let mut event_filter_mask = EventFilter::empty(); | |
563 | ||
923072b8 | 564 | if let Some(event_filters) = event_filters { |
48663c56 XL |
565 | let mut unknown_events = vec![]; |
566 | for item in event_filters { | |
dfeec247 XL |
567 | if let Some(&(_, mask)) = |
568 | EVENT_FILTERS_BY_NAME.iter().find(|&(name, _)| name == item) | |
569 | { | |
48663c56 XL |
570 | event_filter_mask |= mask; |
571 | } else { | |
572 | unknown_events.push(item.clone()); | |
573 | } | |
574 | } | |
575 | ||
576 | // Warn about any unknown event names | |
74b04a01 | 577 | if !unknown_events.is_empty() { |
48663c56 XL |
578 | unknown_events.sort(); |
579 | unknown_events.dedup(); | |
580 | ||
dfeec247 XL |
581 | warn!( |
582 | "Unknown self-profiler events specified: {}. Available options are: {}.", | |
48663c56 | 583 | unknown_events.join(", "), |
dfeec247 XL |
584 | EVENT_FILTERS_BY_NAME |
585 | .iter() | |
586 | .map(|&(name, _)| name.to_string()) | |
587 | .collect::<Vec<_>>() | |
588 | .join(", ") | |
589 | ); | |
48663c56 XL |
590 | } |
591 | } else { | |
592 | event_filter_mask = EventFilter::DEFAULT; | |
593 | } | |
594 | ||
595 | Ok(SelfProfiler { | |
596 | profiler, | |
597 | event_filter_mask, | |
dfeec247 | 598 | string_cache: RwLock::new(FxHashMap::default()), |
48663c56 XL |
599 | query_event_kind, |
600 | generic_activity_event_kind, | |
601 | incremental_load_result_event_kind, | |
136023e0 | 602 | incremental_result_hashing_event_kind, |
48663c56 XL |
603 | query_blocked_event_kind, |
604 | query_cache_hit_event_kind, | |
3c0e092e | 605 | artifact_size_event_kind, |
48663c56 XL |
606 | }) |
607 | } | |
608 | ||
dfeec247 XL |
609 | /// Allocates a new string in the profiling data. Does not do any caching |
610 | /// or deduplication. | |
611 | pub fn alloc_string<STR: SerializableString + ?Sized>(&self, s: &STR) -> StringId { | |
612 | self.profiler.alloc_string(s) | |
613 | } | |
614 | ||
615 | /// Gets a `StringId` for the given string. This method makes sure that | |
616 | /// any strings going through it will only be allocated once in the | |
617 | /// profiling data. | |
74b04a01 XL |
618 | pub fn get_or_alloc_cached_string<A>(&self, s: A) -> StringId |
619 | where | |
620 | A: Borrow<str> + Into<String>, | |
621 | { | |
dfeec247 XL |
622 | // Only acquire a read-lock first since we assume that the string is |
623 | // already present in the common case. | |
624 | { | |
625 | let string_cache = self.string_cache.read(); | |
626 | ||
74b04a01 | 627 | if let Some(&id) = string_cache.get(s.borrow()) { |
dfeec247 XL |
628 | return id; |
629 | } | |
630 | } | |
631 | ||
632 | let mut string_cache = self.string_cache.write(); | |
633 | // Check if the string has already been added in the small time window | |
634 | // between dropping the read lock and acquiring the write lock. | |
74b04a01 XL |
635 | match string_cache.entry(s.into()) { |
636 | Entry::Occupied(e) => *e.get(), | |
637 | Entry::Vacant(e) => { | |
638 | let string_id = self.profiler.alloc_string(&e.key()[..]); | |
639 | *e.insert(string_id) | |
640 | } | |
641 | } | |
dfeec247 XL |
642 | } |
643 | ||
644 | pub fn map_query_invocation_id_to_string(&self, from: QueryInvocationId, to: StringId) { | |
645 | let from = StringId::new_virtual(from.0); | |
646 | self.profiler.map_virtual_to_concrete_string(from, to); | |
647 | } | |
b7449926 | 648 | |
dfeec247 XL |
649 | pub fn bulk_map_query_invocation_id_to_single_string<I>(&self, from: I, to: StringId) |
650 | where | |
651 | I: Iterator<Item = QueryInvocationId> + ExactSizeIterator, | |
652 | { | |
653 | let from = from.map(|qid| StringId::new_virtual(qid.0)); | |
654 | self.profiler.bulk_map_virtual_to_single_concrete_string(from, to); | |
48663c56 XL |
655 | } |
656 | ||
dfeec247 XL |
657 | pub fn query_key_recording_enabled(&self) -> bool { |
658 | self.event_filter_mask.contains(EventFilter::QUERY_KEYS) | |
659 | } | |
660 | ||
29967ef6 | 661 | pub fn event_id_builder(&self) -> EventIdBuilder<'_> { |
dfeec247 | 662 | EventIdBuilder::new(&self.profiler) |
b7449926 | 663 | } |
e74abb32 | 664 | } |
b7449926 | 665 | |
e74abb32 | 666 | #[must_use] |
29967ef6 | 667 | pub struct TimingGuard<'a>(Option<measureme::TimingGuard<'a>>); |
9fa01778 | 668 | |
e74abb32 | 669 | impl<'a> TimingGuard<'a> { |
9fa01778 | 670 | #[inline] |
e74abb32 XL |
671 | pub fn start( |
672 | profiler: &'a SelfProfiler, | |
673 | event_kind: StringId, | |
dfeec247 | 674 | event_id: EventId, |
e74abb32 | 675 | ) -> TimingGuard<'a> { |
3c0e092e | 676 | let thread_id = get_thread_id(); |
e74abb32 | 677 | let raw_profiler = &profiler.profiler; |
dfeec247 XL |
678 | let timing_guard = |
679 | raw_profiler.start_recording_interval_event(event_kind, event_id, thread_id); | |
e74abb32 | 680 | TimingGuard(Some(timing_guard)) |
9fa01778 XL |
681 | } |
682 | ||
dfeec247 XL |
683 | #[inline] |
684 | pub fn finish_with_query_invocation_id(self, query_invocation_id: QueryInvocationId) { | |
685 | if let Some(guard) = self.0 { | |
74b04a01 XL |
686 | cold_path(|| { |
687 | let event_id = StringId::new_virtual(query_invocation_id.0); | |
688 | let event_id = EventId::from_virtual(event_id); | |
689 | guard.finish_with_override_event_id(event_id); | |
690 | }); | |
dfeec247 XL |
691 | } |
692 | } | |
693 | ||
532ac7d7 | 694 | #[inline] |
e74abb32 XL |
695 | pub fn none() -> TimingGuard<'a> { |
696 | TimingGuard(None) | |
b7449926 | 697 | } |
dfeec247 XL |
698 | |
699 | #[inline(always)] | |
700 | pub fn run<R>(self, f: impl FnOnce() -> R) -> R { | |
701 | let _timer = self; | |
702 | f() | |
703 | } | |
704 | } | |
705 | ||
706 | #[must_use] | |
707 | pub struct VerboseTimingGuard<'a> { | |
5869c6ff | 708 | start_and_message: Option<(Instant, Option<usize>, String)>, |
dfeec247 XL |
709 | _guard: TimingGuard<'a>, |
710 | } | |
711 | ||
712 | impl<'a> VerboseTimingGuard<'a> { | |
74b04a01 | 713 | pub fn start(message: Option<String>, _guard: TimingGuard<'a>) -> Self { |
5869c6ff XL |
714 | VerboseTimingGuard { |
715 | _guard, | |
716 | start_and_message: message.map(|msg| (Instant::now(), get_resident_set_size(), msg)), | |
717 | } | |
dfeec247 XL |
718 | } |
719 | ||
720 | #[inline(always)] | |
721 | pub fn run<R>(self, f: impl FnOnce() -> R) -> R { | |
722 | let _timer = self; | |
723 | f() | |
724 | } | |
725 | } | |
726 | ||
727 | impl Drop for VerboseTimingGuard<'_> { | |
728 | fn drop(&mut self) { | |
5869c6ff XL |
729 | if let Some((start_time, start_rss, ref message)) = self.start_and_message { |
730 | let end_rss = get_resident_set_size(); | |
04454e1e | 731 | let dur = start_time.elapsed(); |
487cf647 | 732 | print_time_passes_entry(message, dur, start_rss, end_rss); |
74b04a01 | 733 | } |
dfeec247 XL |
734 | } |
735 | } | |
736 | ||
5869c6ff XL |
737 | pub fn print_time_passes_entry( |
738 | what: &str, | |
739 | dur: Duration, | |
740 | start_rss: Option<usize>, | |
741 | end_rss: Option<usize>, | |
742 | ) { | |
2b03887a FG |
743 | // Print the pass if its duration is greater than 5 ms, or it changed the |
744 | // measured RSS. | |
745 | let is_notable = || { | |
746 | if dur.as_millis() > 5 { | |
747 | return true; | |
748 | } | |
749 | ||
750 | if let (Some(start_rss), Some(end_rss)) = (start_rss, end_rss) { | |
751 | let change_rss = end_rss.abs_diff(start_rss); | |
752 | if change_rss > 0 { | |
753 | return true; | |
754 | } | |
755 | } | |
756 | ||
757 | false | |
758 | }; | |
759 | if !is_notable() { | |
760 | return; | |
761 | } | |
762 | ||
5869c6ff XL |
763 | let rss_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as usize; |
764 | let rss_change_to_mb = |rss| (rss as f64 / 1_000_000.0).round() as i128; | |
765 | ||
766 | let mem_string = match (start_rss, end_rss) { | |
767 | (Some(start_rss), Some(end_rss)) => { | |
768 | let change_rss = end_rss as i128 - start_rss as i128; | |
769 | ||
770 | format!( | |
771 | "; rss: {:>4}MB -> {:>4}MB ({:>+5}MB)", | |
772 | rss_to_mb(start_rss), | |
773 | rss_to_mb(end_rss), | |
774 | rss_change_to_mb(change_rss), | |
775 | ) | |
dfeec247 | 776 | } |
5869c6ff XL |
777 | (Some(start_rss), None) => format!("; rss start: {:>4}MB", rss_to_mb(start_rss)), |
778 | (None, Some(end_rss)) => format!("; rss end: {:>4}MB", rss_to_mb(end_rss)), | |
779 | (None, None) => String::new(), | |
dfeec247 | 780 | }; |
5869c6ff | 781 | |
6a06907d | 782 | eprintln!("time: {:>7}{}\t{}", duration_to_secs_str(dur), mem_string, what); |
dfeec247 XL |
783 | } |
784 | ||
785 | // Hack up our own formatting for the duration to make it easier for scripts | |
786 | // to parse (always use the same number of decimal places and the same unit). | |
787 | pub fn duration_to_secs_str(dur: std::time::Duration) -> String { | |
1b1a35ee | 788 | format!("{:.3}", dur.as_secs_f64()) |
dfeec247 XL |
789 | } |
790 | ||
3c0e092e XL |
791 | fn get_thread_id() -> u32 { |
792 | std::thread::current().id().as_u64().get() as u32 | |
793 | } | |
794 | ||
dfeec247 | 795 | // Memory reporting |
f9f354fc XL |
796 | cfg_if! { |
797 | if #[cfg(windows)] { | |
5869c6ff | 798 | pub fn get_resident_set_size() -> Option<usize> { |
f9f354fc XL |
799 | use std::mem::{self, MaybeUninit}; |
800 | use winapi::shared::minwindef::DWORD; | |
801 | use winapi::um::processthreadsapi::GetCurrentProcess; | |
802 | use winapi::um::psapi::{GetProcessMemoryInfo, PROCESS_MEMORY_COUNTERS}; | |
803 | ||
804 | let mut pmc = MaybeUninit::<PROCESS_MEMORY_COUNTERS>::uninit(); | |
805 | match unsafe { | |
806 | GetProcessMemoryInfo(GetCurrentProcess(), pmc.as_mut_ptr(), mem::size_of_val(&pmc) as DWORD) | |
807 | } { | |
808 | 0 => None, | |
809 | _ => { | |
810 | let pmc = unsafe { pmc.assume_init() }; | |
811 | Some(pmc.WorkingSetSize as usize) | |
812 | } | |
813 | } | |
814 | } | |
064997fb FG |
815 | } else if #[cfg(target_os = "macos")] { |
816 | pub fn get_resident_set_size() -> Option<usize> { | |
817 | use libc::{c_int, c_void, getpid, proc_pidinfo, proc_taskinfo, PROC_PIDTASKINFO}; | |
818 | use std::mem; | |
819 | const PROC_TASKINFO_SIZE: c_int = mem::size_of::<proc_taskinfo>() as c_int; | |
820 | ||
821 | unsafe { | |
822 | let mut info: proc_taskinfo = mem::zeroed(); | |
823 | let info_ptr = &mut info as *mut proc_taskinfo as *mut c_void; | |
824 | let pid = getpid() as c_int; | |
825 | let ret = proc_pidinfo(pid, PROC_PIDTASKINFO, 0, info_ptr, PROC_TASKINFO_SIZE); | |
826 | if ret == PROC_TASKINFO_SIZE { | |
827 | Some(info.pti_resident_size as usize) | |
828 | } else { | |
829 | None | |
830 | } | |
831 | } | |
832 | } | |
f9f354fc | 833 | } else if #[cfg(unix)] { |
5869c6ff | 834 | pub fn get_resident_set_size() -> Option<usize> { |
f9f354fc XL |
835 | let field = 1; |
836 | let contents = fs::read("/proc/self/statm").ok()?; | |
837 | let contents = String::from_utf8(contents).ok()?; | |
838 | let s = contents.split_whitespace().nth(field)?; | |
839 | let npages = s.parse::<usize>().ok()?; | |
840 | Some(npages * 4096) | |
841 | } | |
842 | } else { | |
5869c6ff | 843 | pub fn get_resident_set_size() -> Option<usize> { |
f9f354fc | 844 | None |
dfeec247 XL |
845 | } |
846 | } | |
b7449926 | 847 | } |