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}
;
98 use measureme
::{EventId, EventIdBuilder, SerializableString, StringId}
;
99 use parking_lot
::RwLock
;
101 /// MmapSerializatioSink is faster on macOS and Linux
102 /// but FileSerializationSink is faster on Windows
104 type SerializationSink
= measureme
::MmapSerializationSink
;
106 type SerializationSink
= measureme
::FileSerializationSink
;
108 type Profiler
= measureme
::Profiler
<SerializationSink
>;
110 #[derive(Clone, Copy, Debug, PartialEq, Eq, Ord, PartialOrd)]
111 pub enum ProfileCategory
{
121 bitflags
::bitflags
! {
122 struct EventFilter
: u32 {
123 const GENERIC_ACTIVITIES
= 1 << 0;
124 const QUERY_PROVIDERS
= 1 << 1;
125 const QUERY_CACHE_HITS
= 1 << 2;
126 const QUERY_BLOCKED
= 1 << 3;
127 const INCR_CACHE_LOADS
= 1 << 4;
129 const QUERY_KEYS
= 1 << 5;
130 const FUNCTION_ARGS
= 1 << 6;
133 const DEFAULT
= Self::GENERIC_ACTIVITIES
.bits
|
134 Self::QUERY_PROVIDERS
.bits
|
135 Self::QUERY_BLOCKED
.bits
|
136 Self::INCR_CACHE_LOADS
.bits
;
138 const ARGS
= Self::QUERY_KEYS
.bits
| Self::FUNCTION_ARGS
.bits
;
142 // keep this in sync with the `-Z self-profile-events` help message in librustc_session/options.rs
143 const EVENT_FILTERS_BY_NAME
: &[(&str, EventFilter
)] = &[
144 ("none", EventFilter
::empty()),
145 ("all", EventFilter
::all()),
146 ("default", EventFilter
::DEFAULT
),
147 ("generic-activity", EventFilter
::GENERIC_ACTIVITIES
),
148 ("query-provider", EventFilter
::QUERY_PROVIDERS
),
149 ("query-cache-hit", EventFilter
::QUERY_CACHE_HITS
),
150 ("query-blocked", EventFilter
::QUERY_BLOCKED
),
151 ("incr-cache-load", EventFilter
::INCR_CACHE_LOADS
),
152 ("query-keys", EventFilter
::QUERY_KEYS
),
153 ("function-args", EventFilter
::FUNCTION_ARGS
),
154 ("args", EventFilter
::ARGS
),
155 ("llvm", EventFilter
::LLVM
),
158 /// Something that uniquely identifies a query invocation.
159 pub struct QueryInvocationId(pub u32);
161 /// A reference to the SelfProfiler. It can be cloned and sent across thread
162 /// boundaries at will.
164 pub struct SelfProfilerRef
{
165 // This field is `None` if self-profiling is disabled for the current
166 // compilation session.
167 profiler
: Option
<Arc
<SelfProfiler
>>,
169 // We store the filter mask directly in the reference because that doesn't
170 // cost anything and allows for filtering with checking if the profiler is
172 event_filter_mask
: EventFilter
,
174 // Print verbose generic activities to stdout
175 print_verbose_generic_activities
: bool
,
177 // Print extra verbose generic activities to stdout
178 print_extra_verbose_generic_activities
: bool
,
181 impl SelfProfilerRef
{
183 profiler
: Option
<Arc
<SelfProfiler
>>,
184 print_verbose_generic_activities
: bool
,
185 print_extra_verbose_generic_activities
: bool
,
186 ) -> SelfProfilerRef
{
187 // If there is no SelfProfiler then the filter mask is set to NONE,
188 // ensuring that nothing ever tries to actually access it.
189 let event_filter_mask
=
190 profiler
.as_ref().map(|p
| p
.event_filter_mask
).unwrap_or(EventFilter
::empty());
195 print_verbose_generic_activities
,
196 print_extra_verbose_generic_activities
,
200 // This shim makes sure that calls only get executed if the filter mask
201 // lets them pass. It also contains some trickery to make sure that
202 // code is optimized for non-profiling compilation sessions, i.e. anything
203 // past the filter check is never inlined so it doesn't clutter the fast
206 fn exec
<F
>(&self, event_filter
: EventFilter
, f
: F
) -> TimingGuard
<'_
>
208 F
: for<'a
> FnOnce(&'a SelfProfiler
) -> TimingGuard
<'a
>,
211 fn cold_call
<F
>(profiler_ref
: &SelfProfilerRef
, f
: F
) -> TimingGuard
<'_
>
213 F
: for<'a
> FnOnce(&'a SelfProfiler
) -> TimingGuard
<'a
>,
215 let profiler
= profiler_ref
.profiler
.as_ref().unwrap();
219 if unlikely
!(self.event_filter_mask
.contains(event_filter
)) {
226 /// Start profiling a verbose generic activity. Profiling continues until the
227 /// VerboseTimingGuard returned from this call is dropped. In addition to recording
228 /// a measureme event, "verbose" generic activities also print a timing entry to
229 /// stdout if the compiler is invoked with -Ztime or -Ztime-passes.
230 pub fn verbose_generic_activity
<'a
>(
232 event_label
: &'
static str,
233 ) -> VerboseTimingGuard
<'a
> {
235 if self.print_verbose_generic_activities { Some(event_label.to_owned()) }
else { None }
;
237 VerboseTimingGuard
::start(message
, self.generic_activity(event_label
))
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.
244 pub fn extra_verbose_generic_activity
<'a
, A
>(
246 event_label
: &'
static str,
248 ) -> VerboseTimingGuard
<'a
>
250 A
: Borrow
<str> + Into
<String
>,
252 let message
= if self.print_extra_verbose_generic_activities
{
253 Some(format
!("{}({})", event_label
, event_arg
.borrow()))
258 VerboseTimingGuard
::start(message
, self.generic_activity_with_arg(event_label
, event_arg
))
261 /// Start profiling a generic activity. Profiling continues until the
262 /// TimingGuard returned from this call is dropped.
264 pub fn generic_activity(&self, event_label
: &'
static str) -> TimingGuard
<'_
> {
265 self.exec(EventFilter
::GENERIC_ACTIVITIES
, |profiler
| {
266 let event_label
= profiler
.get_or_alloc_cached_string(event_label
);
267 let event_id
= EventId
::from_label(event_label
);
268 TimingGuard
::start(profiler
, profiler
.generic_activity_event_kind
, event_id
)
272 /// Start profiling a generic activity. Profiling continues until the
273 /// TimingGuard returned from this call is dropped.
275 pub fn generic_activity_with_arg
<A
>(
277 event_label
: &'
static str,
281 A
: Borrow
<str> + Into
<String
>,
283 self.exec(EventFilter
::GENERIC_ACTIVITIES
, |profiler
| {
284 let builder
= EventIdBuilder
::new(&profiler
.profiler
);
285 let event_label
= profiler
.get_or_alloc_cached_string(event_label
);
286 let event_id
= if profiler
.event_filter_mask
.contains(EventFilter
::FUNCTION_ARGS
) {
287 let event_arg
= profiler
.get_or_alloc_cached_string(event_arg
);
288 builder
.from_label_and_arg(event_label
, event_arg
)
290 builder
.from_label(event_label
)
292 TimingGuard
::start(profiler
, profiler
.generic_activity_event_kind
, event_id
)
296 /// Start profiling a query provider. Profiling continues until the
297 /// TimingGuard returned from this call is dropped.
299 pub fn query_provider(&self) -> TimingGuard
<'_
> {
300 self.exec(EventFilter
::QUERY_PROVIDERS
, |profiler
| {
301 TimingGuard
::start(profiler
, profiler
.query_event_kind
, EventId
::INVALID
)
305 /// Record a query in-memory cache hit.
307 pub fn query_cache_hit(&self, query_invocation_id
: QueryInvocationId
) {
308 self.instant_query_event(
309 |profiler
| profiler
.query_cache_hit_event_kind
,
311 EventFilter
::QUERY_CACHE_HITS
,
315 /// Start profiling a query being blocked on a concurrent execution.
316 /// Profiling continues until the TimingGuard returned from this call is
319 pub fn query_blocked(&self) -> TimingGuard
<'_
> {
320 self.exec(EventFilter
::QUERY_BLOCKED
, |profiler
| {
321 TimingGuard
::start(profiler
, profiler
.query_blocked_event_kind
, EventId
::INVALID
)
325 /// Start profiling how long it takes to load a query result from the
326 /// incremental compilation on-disk cache. Profiling continues until the
327 /// TimingGuard returned from this call is dropped.
329 pub fn incr_cache_loading(&self) -> TimingGuard
<'_
> {
330 self.exec(EventFilter
::INCR_CACHE_LOADS
, |profiler
| {
333 profiler
.incremental_load_result_event_kind
,
340 fn instant_query_event(
342 event_kind
: fn(&SelfProfiler
) -> StringId
,
343 query_invocation_id
: QueryInvocationId
,
344 event_filter
: EventFilter
,
346 drop(self.exec(event_filter
, |profiler
| {
347 let event_id
= StringId
::new_virtual(query_invocation_id
.0);
348 let thread_id
= std
::thread
::current().id().as_u64() as u32;
350 profiler
.profiler
.record_instant_event(
351 event_kind(profiler
),
352 EventId
::from_virtual(event_id
),
360 pub fn with_profiler(&self, f
: impl FnOnce(&SelfProfiler
)) {
361 if let Some(profiler
) = &self.profiler
{
367 pub fn enabled(&self) -> bool
{
368 self.profiler
.is_some()
372 pub fn llvm_recording_enabled(&self) -> bool
{
373 self.event_filter_mask
.contains(EventFilter
::LLVM
)
376 pub fn get_self_profiler(&self) -> Option
<Arc
<SelfProfiler
>> {
377 self.profiler
.clone()
381 pub struct SelfProfiler
{
383 event_filter_mask
: EventFilter
,
385 string_cache
: RwLock
<FxHashMap
<String
, StringId
>>,
387 query_event_kind
: StringId
,
388 generic_activity_event_kind
: StringId
,
389 incremental_load_result_event_kind
: StringId
,
390 query_blocked_event_kind
: StringId
,
391 query_cache_hit_event_kind
: StringId
,
396 output_directory
: &Path
,
397 crate_name
: Option
<&str>,
398 event_filters
: &Option
<Vec
<String
>>,
399 ) -> Result
<SelfProfiler
, Box
<dyn Error
>> {
400 fs
::create_dir_all(output_directory
)?
;
402 let crate_name
= crate_name
.unwrap_or("unknown-crate");
403 let filename
= format
!("{}-{}.rustc_profile", crate_name
, process
::id());
404 let path
= output_directory
.join(&filename
);
405 let profiler
= Profiler
::new(&path
)?
;
407 let query_event_kind
= profiler
.alloc_string("Query");
408 let generic_activity_event_kind
= profiler
.alloc_string("GenericActivity");
409 let incremental_load_result_event_kind
= profiler
.alloc_string("IncrementalLoadResult");
410 let query_blocked_event_kind
= profiler
.alloc_string("QueryBlocked");
411 let query_cache_hit_event_kind
= profiler
.alloc_string("QueryCacheHit");
413 let mut event_filter_mask
= EventFilter
::empty();
415 if let Some(ref event_filters
) = *event_filters
{
416 let mut unknown_events
= vec
![];
417 for item
in event_filters
{
418 if let Some(&(_
, mask
)) =
419 EVENT_FILTERS_BY_NAME
.iter().find(|&(name
, _
)| name
== item
)
421 event_filter_mask
|= mask
;
423 unknown_events
.push(item
.clone());
427 // Warn about any unknown event names
428 if !unknown_events
.is_empty() {
429 unknown_events
.sort();
430 unknown_events
.dedup();
433 "Unknown self-profiler events specified: {}. Available options are: {}.",
434 unknown_events
.join(", "),
435 EVENT_FILTERS_BY_NAME
437 .map(|&(name
, _
)| name
.to_string())
443 event_filter_mask
= EventFilter
::DEFAULT
;
449 string_cache
: RwLock
::new(FxHashMap
::default()),
451 generic_activity_event_kind
,
452 incremental_load_result_event_kind
,
453 query_blocked_event_kind
,
454 query_cache_hit_event_kind
,
458 /// Allocates a new string in the profiling data. Does not do any caching
459 /// or deduplication.
460 pub fn alloc_string
<STR
: SerializableString
+ ?Sized
>(&self, s
: &STR
) -> StringId
{
461 self.profiler
.alloc_string(s
)
464 /// Gets a `StringId` for the given string. This method makes sure that
465 /// any strings going through it will only be allocated once in the
467 pub fn get_or_alloc_cached_string
<A
>(&self, s
: A
) -> StringId
469 A
: Borrow
<str> + Into
<String
>,
471 // Only acquire a read-lock first since we assume that the string is
472 // already present in the common case.
474 let string_cache
= self.string_cache
.read();
476 if let Some(&id
) = string_cache
.get(s
.borrow()) {
481 let mut string_cache
= self.string_cache
.write();
482 // Check if the string has already been added in the small time window
483 // between dropping the read lock and acquiring the write lock.
484 match string_cache
.entry(s
.into()) {
485 Entry
::Occupied(e
) => *e
.get(),
486 Entry
::Vacant(e
) => {
487 let string_id
= self.profiler
.alloc_string(&e
.key()[..]);
493 pub fn map_query_invocation_id_to_string(&self, from
: QueryInvocationId
, to
: StringId
) {
494 let from
= StringId
::new_virtual(from
.0);
495 self.profiler
.map_virtual_to_concrete_string(from
, to
);
498 pub fn bulk_map_query_invocation_id_to_single_string
<I
>(&self, from
: I
, to
: StringId
)
500 I
: Iterator
<Item
= QueryInvocationId
> + ExactSizeIterator
,
502 let from
= from
.map(|qid
| StringId
::new_virtual(qid
.0));
503 self.profiler
.bulk_map_virtual_to_single_concrete_string(from
, to
);
506 pub fn query_key_recording_enabled(&self) -> bool
{
507 self.event_filter_mask
.contains(EventFilter
::QUERY_KEYS
)
510 pub fn event_id_builder(&self) -> EventIdBuilder
<'_
, SerializationSink
> {
511 EventIdBuilder
::new(&self.profiler
)
516 pub struct TimingGuard
<'a
>(Option
<measureme
::TimingGuard
<'a
, SerializationSink
>>);
518 impl<'a
> TimingGuard
<'a
> {
521 profiler
: &'a SelfProfiler
,
522 event_kind
: StringId
,
524 ) -> TimingGuard
<'a
> {
525 let thread_id
= std
::thread
::current().id().as_u64() as u32;
526 let raw_profiler
= &profiler
.profiler
;
528 raw_profiler
.start_recording_interval_event(event_kind
, event_id
, thread_id
);
529 TimingGuard(Some(timing_guard
))
533 pub fn finish_with_query_invocation_id(self, query_invocation_id
: QueryInvocationId
) {
534 if let Some(guard
) = self.0 {
536 let event_id
= StringId
::new_virtual(query_invocation_id
.0);
537 let event_id
= EventId
::from_virtual(event_id
);
538 guard
.finish_with_override_event_id(event_id
);
544 pub fn none() -> TimingGuard
<'a
> {
549 pub fn run
<R
>(self, f
: impl FnOnce() -> R
) -> R
{
556 pub struct VerboseTimingGuard
<'a
> {
557 start_and_message
: Option
<(Instant
, String
)>,
558 _guard
: TimingGuard
<'a
>,
561 impl<'a
> VerboseTimingGuard
<'a
> {
562 pub fn start(message
: Option
<String
>, _guard
: TimingGuard
<'a
>) -> Self {
563 VerboseTimingGuard { _guard, start_and_message: message.map(|msg| (Instant::now(), msg)) }
567 pub fn run
<R
>(self, f
: impl FnOnce() -> R
) -> R
{
573 impl Drop
for VerboseTimingGuard
<'_
> {
575 if let Some((start
, ref message
)) = self.start_and_message
{
576 print_time_passes_entry(true, &message
[..], start
.elapsed());
581 pub fn print_time_passes_entry(do_it
: bool
, what
: &str, dur
: Duration
) {
586 let mem_string
= match get_resident() {
588 let mb
= n
as f64 / 1_000_000.0;
589 format
!("; rss: {}MB", mb
.round() as usize)
591 None
=> String
::new(),
593 println
!("time: {}{}\t{}", duration_to_secs_str(dur
), mem_string
, what
);
596 // Hack up our own formatting for the duration to make it easier for scripts
597 // to parse (always use the same number of decimal places and the same unit).
598 pub fn duration_to_secs_str(dur
: std
::time
::Duration
) -> String
{
599 const NANOS_PER_SEC
: f64 = 1_000_000_000.0;
600 let secs
= dur
.as_secs() as f64 + dur
.subsec_nanos() as f64 / NANOS_PER_SEC
;
602 format
!("{:.3}", secs
)
607 fn get_resident() -> Option
<usize> {
609 let contents
= fs
::read("/proc/self/statm").ok()?
;
610 let contents
= String
::from_utf8(contents
).ok()?
;
611 let s
= contents
.split_whitespace().nth(field
)?
;
612 let npages
= s
.parse
::<usize>().ok()?
;
617 fn get_resident() -> Option
<usize> {
618 use std
::mem
::{self, MaybeUninit}
;
619 use winapi
::shared
::minwindef
::DWORD
;
620 use winapi
::um
::processthreadsapi
::GetCurrentProcess
;
621 use winapi
::um
::psapi
::{GetProcessMemoryInfo, PROCESS_MEMORY_COUNTERS}
;
623 let mut pmc
= MaybeUninit
::<PROCESS_MEMORY_COUNTERS
>::uninit();
625 GetProcessMemoryInfo(GetCurrentProcess(), pmc
.as_mut_ptr(), mem
::size_of_val(&pmc
) as DWORD
)
629 let pmc
= unsafe { pmc.assume_init() }
;
630 Some(pmc
.WorkingSetSize
as usize)