1 | //===-- PerfReader.h - perfscript reader -----------------------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #ifndef LLVM_TOOLS_LLVM_PROFGEN_PERFREADER_H |
10 | #define LLVM_TOOLS_LLVM_PROFGEN_PERFREADER_H |
11 | #include "ErrorHandling.h" |
12 | #include "ProfiledBinary.h" |
13 | #include "llvm/Support/Casting.h" |
14 | #include "llvm/Support/CommandLine.h" |
15 | #include "llvm/Support/Regex.h" |
16 | #include <cstdint> |
17 | #include <fstream> |
18 | #include <map> |
19 | |
20 | using namespace llvm; |
21 | using namespace sampleprof; |
22 | |
23 | namespace llvm { |
24 | |
25 | class CleanupInstaller; |
26 | |
27 | namespace sampleprof { |
28 | |
29 | // Stream based trace line iterator |
30 | class TraceStream { |
31 | std::string CurrentLine; |
32 | std::ifstream Fin; |
33 | bool IsAtEoF = false; |
34 | uint64_t LineNumber = 0; |
35 | |
36 | public: |
37 | TraceStream(StringRef Filename) : Fin(Filename.str()) { |
38 | if (!Fin.good()) |
39 | exitWithError(Message: "Error read input perf script file" , Whence: Filename); |
40 | advance(); |
41 | } |
42 | |
43 | StringRef getCurrentLine() { |
44 | assert(!IsAtEoF && "Line iterator reaches the End-of-File!" ); |
45 | return CurrentLine; |
46 | } |
47 | |
48 | uint64_t getLineNumber() { return LineNumber; } |
49 | |
50 | bool isAtEoF() { return IsAtEoF; } |
51 | |
52 | // Read the next line |
53 | void advance() { |
54 | if (!std::getline(is&: Fin, str&: CurrentLine)) { |
55 | IsAtEoF = true; |
56 | return; |
57 | } |
58 | LineNumber++; |
59 | } |
60 | }; |
61 | |
62 | // The type of input format. |
63 | enum PerfFormat { |
64 | UnknownFormat = 0, |
65 | PerfData = 1, // Raw linux perf.data. |
66 | PerfScript = 2, // Perf script create by `perf script` command. |
67 | UnsymbolizedProfile = 3, // Unsymbolized profile generated by llvm-profgen. |
68 | |
69 | }; |
70 | |
71 | // The type of perfscript content. |
72 | enum PerfContent { |
73 | UnknownContent = 0, |
74 | LBR = 1, // Only LBR sample. |
75 | LBRStack = 2, // Hybrid sample including call stack and LBR stack. |
76 | }; |
77 | |
78 | struct PerfInputFile { |
79 | std::string InputFile; |
80 | PerfFormat Format = PerfFormat::UnknownFormat; |
81 | PerfContent Content = PerfContent::UnknownContent; |
82 | }; |
83 | |
84 | // The parsed LBR sample entry. |
85 | struct LBREntry { |
86 | uint64_t Source = 0; |
87 | uint64_t Target = 0; |
88 | LBREntry(uint64_t S, uint64_t T) : Source(S), Target(T) {} |
89 | |
90 | #ifndef NDEBUG |
91 | void print() const { |
92 | dbgs() << "from " << format("%#010x" , Source) << " to " |
93 | << format("%#010x" , Target); |
94 | } |
95 | #endif |
96 | }; |
97 | |
98 | #ifndef NDEBUG |
99 | static inline void printLBRStack(const SmallVectorImpl<LBREntry> &LBRStack) { |
100 | for (size_t I = 0; I < LBRStack.size(); I++) { |
101 | dbgs() << "[" << I << "] " ; |
102 | LBRStack[I].print(); |
103 | dbgs() << "\n" ; |
104 | } |
105 | } |
106 | |
107 | static inline void printCallStack(const SmallVectorImpl<uint64_t> &CallStack) { |
108 | for (size_t I = 0; I < CallStack.size(); I++) { |
109 | dbgs() << "[" << I << "] " << format("%#010x" , CallStack[I]) << "\n" ; |
110 | } |
111 | } |
112 | #endif |
113 | |
114 | // Hash interface for generic data of type T |
115 | // Data should implement a \fn getHashCode and a \fn isEqual |
116 | // Currently getHashCode is non-virtual to avoid the overhead of calling vtable, |
117 | // i.e we explicitly calculate hash of derived class, assign to base class's |
118 | // HashCode. This also provides the flexibility for calculating the hash code |
119 | // incrementally(like rolling hash) during frame stack unwinding since unwinding |
120 | // only changes the leaf of frame stack. \fn isEqual is a virtual function, |
121 | // which will have perf overhead. In the future, if we redesign a better hash |
122 | // function, then we can just skip this or switch to non-virtual function(like |
123 | // just ignore comparison if hash conflicts probabilities is low) |
124 | template <class T> class Hashable { |
125 | public: |
126 | std::shared_ptr<T> Data; |
127 | Hashable(const std::shared_ptr<T> &D) : Data(D) {} |
128 | |
129 | // Hash code generation |
130 | struct Hash { |
131 | uint64_t operator()(const Hashable<T> &Key) const { |
132 | // Don't make it virtual for getHashCode |
133 | uint64_t Hash = Key.Data->getHashCode(); |
134 | assert(Hash && "Should generate HashCode for it!" ); |
135 | return Hash; |
136 | } |
137 | }; |
138 | |
139 | // Hash equal |
140 | struct Equal { |
141 | bool operator()(const Hashable<T> &LHS, const Hashable<T> &RHS) const { |
142 | // Precisely compare the data, vtable will have overhead. |
143 | return LHS.Data->isEqual(RHS.Data.get()); |
144 | } |
145 | }; |
146 | |
147 | T *getPtr() const { return Data.get(); } |
148 | }; |
149 | |
150 | struct PerfSample { |
151 | // LBR stack recorded in FIFO order. |
152 | SmallVector<LBREntry, 16> LBRStack; |
153 | // Call stack recorded in FILO(leaf to root) order, it's used for CS-profile |
154 | // generation |
155 | SmallVector<uint64_t, 16> CallStack; |
156 | |
157 | virtual ~PerfSample() = default; |
158 | uint64_t getHashCode() const { |
159 | // Use simple DJB2 hash |
160 | auto HashCombine = [](uint64_t H, uint64_t V) { |
161 | return ((H << 5) + H) + V; |
162 | }; |
163 | uint64_t Hash = 5381; |
164 | for (const auto &Value : CallStack) { |
165 | Hash = HashCombine(Hash, Value); |
166 | } |
167 | for (const auto &Entry : LBRStack) { |
168 | Hash = HashCombine(Hash, Entry.Source); |
169 | Hash = HashCombine(Hash, Entry.Target); |
170 | } |
171 | return Hash; |
172 | } |
173 | |
174 | bool isEqual(const PerfSample *Other) const { |
175 | const SmallVector<uint64_t, 16> &OtherCallStack = Other->CallStack; |
176 | const SmallVector<LBREntry, 16> &OtherLBRStack = Other->LBRStack; |
177 | |
178 | if (CallStack.size() != OtherCallStack.size() || |
179 | LBRStack.size() != OtherLBRStack.size()) |
180 | return false; |
181 | |
182 | if (!std::equal(CallStack.begin(), CallStack.end(), OtherCallStack.begin())) |
183 | return false; |
184 | |
185 | for (size_t I = 0; I < OtherLBRStack.size(); I++) { |
186 | if (LBRStack[I].Source != OtherLBRStack[I].Source || |
187 | LBRStack[I].Target != OtherLBRStack[I].Target) |
188 | return false; |
189 | } |
190 | return true; |
191 | } |
192 | |
193 | #ifndef NDEBUG |
194 | uint64_t Linenum = 0; |
195 | |
196 | void print() const { |
197 | dbgs() << "Line " << Linenum << "\n" ; |
198 | dbgs() << "LBR stack\n" ; |
199 | printLBRStack(LBRStack); |
200 | dbgs() << "Call stack\n" ; |
201 | printCallStack(CallStack); |
202 | } |
203 | #endif |
204 | }; |
205 | // After parsing the sample, we record the samples by aggregating them |
206 | // into this counter. The key stores the sample data and the value is |
207 | // the sample repeat times. |
208 | using AggregatedCounter = |
209 | std::unordered_map<Hashable<PerfSample>, uint64_t, |
210 | Hashable<PerfSample>::Hash, Hashable<PerfSample>::Equal>; |
211 | |
212 | using SampleVector = SmallVector<std::tuple<uint64_t, uint64_t, uint64_t>, 16>; |
213 | |
214 | inline bool isValidFallThroughRange(uint64_t Start, uint64_t End, |
215 | ProfiledBinary *Binary) { |
216 | // Start bigger than End is considered invalid. |
217 | // LBR ranges cross the unconditional jmp are also assumed invalid. |
218 | // It's found that perf data may contain duplicate LBR entries that could form |
219 | // a range that does not reflect real execution flow on some Intel targets, |
220 | // e.g. Skylake. Such ranges are ususally very long. Exclude them since there |
221 | // cannot be a linear execution range that spans over unconditional jmp. |
222 | return Start <= End && !Binary->rangeCrossUncondBranch(Start, End); |
223 | } |
224 | |
225 | // The state for the unwinder, it doesn't hold the data but only keep the |
226 | // pointer/index of the data, While unwinding, the CallStack is changed |
227 | // dynamicially and will be recorded as the context of the sample |
228 | struct UnwindState { |
229 | // Profiled binary that current frame address belongs to |
230 | const ProfiledBinary *Binary; |
231 | // Call stack trie node |
232 | struct ProfiledFrame { |
233 | const uint64_t Address = DummyRoot; |
234 | ProfiledFrame *Parent; |
235 | SampleVector RangeSamples; |
236 | SampleVector BranchSamples; |
237 | std::unordered_map<uint64_t, std::unique_ptr<ProfiledFrame>> Children; |
238 | |
239 | ProfiledFrame(uint64_t Addr = 0, ProfiledFrame *P = nullptr) |
240 | : Address(Addr), Parent(P) {} |
241 | ProfiledFrame *getOrCreateChildFrame(uint64_t Address) { |
242 | assert(Address && "Address can't be zero!" ); |
243 | auto Ret = Children.emplace( |
244 | args&: Address, args: std::make_unique<ProfiledFrame>(args&: Address, args: this)); |
245 | return Ret.first->second.get(); |
246 | } |
247 | void recordRangeCount(uint64_t Start, uint64_t End, uint64_t Count) { |
248 | RangeSamples.emplace_back(Args: std::make_tuple(args&: Start, args&: End, args&: Count)); |
249 | } |
250 | void recordBranchCount(uint64_t Source, uint64_t Target, uint64_t Count) { |
251 | BranchSamples.emplace_back(Args: std::make_tuple(args&: Source, args&: Target, args&: Count)); |
252 | } |
253 | bool isDummyRoot() { return Address == DummyRoot; } |
254 | bool isExternalFrame() { return Address == ExternalAddr; } |
255 | bool isLeafFrame() { return Children.empty(); } |
256 | }; |
257 | |
258 | ProfiledFrame DummyTrieRoot; |
259 | ProfiledFrame *CurrentLeafFrame; |
260 | // Used to fall through the LBR stack |
261 | uint32_t LBRIndex = 0; |
262 | // Reference to PerfSample.LBRStack |
263 | const SmallVector<LBREntry, 16> &LBRStack; |
264 | // Used to iterate the address range |
265 | InstructionPointer InstPtr; |
266 | // Indicate whether unwinding is currently in a bad state which requires to |
267 | // skip all subsequent unwinding. |
268 | bool Invalid = false; |
269 | UnwindState(const PerfSample *Sample, const ProfiledBinary *Binary) |
270 | : Binary(Binary), LBRStack(Sample->LBRStack), |
271 | InstPtr(Binary, Sample->CallStack.front()) { |
272 | initFrameTrie(CallStack: Sample->CallStack); |
273 | } |
274 | |
275 | bool validateInitialState() { |
276 | uint64_t LBRLeaf = LBRStack[LBRIndex].Target; |
277 | uint64_t LeafAddr = CurrentLeafFrame->Address; |
278 | assert((LBRLeaf != ExternalAddr || LBRLeaf == LeafAddr) && |
279 | "External leading LBR should match the leaf frame." ); |
280 | |
281 | // When we take a stack sample, ideally the sampling distance between the |
282 | // leaf IP of stack and the last LBR target shouldn't be very large. |
283 | // Use a heuristic size (0x100) to filter out broken records. |
284 | if (LeafAddr < LBRLeaf || LeafAddr - LBRLeaf >= 0x100) { |
285 | WithColor::warning() << "Bogus trace: stack tip = " |
286 | << format(Fmt: "%#010x" , Vals: LeafAddr) |
287 | << ", LBR tip = " << format(Fmt: "%#010x\n" , Vals: LBRLeaf); |
288 | return false; |
289 | } |
290 | return true; |
291 | } |
292 | |
293 | void checkStateConsistency() { |
294 | assert(InstPtr.Address == CurrentLeafFrame->Address && |
295 | "IP should align with context leaf" ); |
296 | } |
297 | |
298 | void setInvalid() { Invalid = true; } |
299 | bool hasNextLBR() const { return LBRIndex < LBRStack.size(); } |
300 | uint64_t getCurrentLBRSource() const { return LBRStack[LBRIndex].Source; } |
301 | uint64_t getCurrentLBRTarget() const { return LBRStack[LBRIndex].Target; } |
302 | const LBREntry &getCurrentLBR() const { return LBRStack[LBRIndex]; } |
303 | bool IsLastLBR() const { return LBRIndex == 0; } |
304 | bool getLBRStackSize() const { return LBRStack.size(); } |
305 | void advanceLBR() { LBRIndex++; } |
306 | ProfiledFrame *getParentFrame() { return CurrentLeafFrame->Parent; } |
307 | |
308 | void pushFrame(uint64_t Address) { |
309 | CurrentLeafFrame = CurrentLeafFrame->getOrCreateChildFrame(Address); |
310 | } |
311 | |
312 | void switchToFrame(uint64_t Address) { |
313 | if (CurrentLeafFrame->Address == Address) |
314 | return; |
315 | CurrentLeafFrame = CurrentLeafFrame->Parent->getOrCreateChildFrame(Address); |
316 | } |
317 | |
318 | void popFrame() { CurrentLeafFrame = CurrentLeafFrame->Parent; } |
319 | |
320 | void clearCallStack() { CurrentLeafFrame = &DummyTrieRoot; } |
321 | |
322 | void initFrameTrie(const SmallVectorImpl<uint64_t> &CallStack) { |
323 | ProfiledFrame *Cur = &DummyTrieRoot; |
324 | for (auto Address : reverse(C: CallStack)) { |
325 | Cur = Cur->getOrCreateChildFrame(Address); |
326 | } |
327 | CurrentLeafFrame = Cur; |
328 | } |
329 | |
330 | ProfiledFrame *getDummyRootPtr() { return &DummyTrieRoot; } |
331 | }; |
332 | |
333 | // Base class for sample counter key with context |
334 | struct ContextKey { |
335 | uint64_t HashCode = 0; |
336 | virtual ~ContextKey() = default; |
337 | uint64_t getHashCode() { |
338 | if (HashCode == 0) |
339 | genHashCode(); |
340 | return HashCode; |
341 | } |
342 | virtual void genHashCode() = 0; |
343 | virtual bool isEqual(const ContextKey *K) const { |
344 | return HashCode == K->HashCode; |
345 | }; |
346 | |
347 | // Utilities for LLVM-style RTTI |
348 | enum ContextKind { CK_StringBased, CK_AddrBased }; |
349 | const ContextKind Kind; |
350 | ContextKind getKind() const { return Kind; } |
351 | ContextKey(ContextKind K) : Kind(K){}; |
352 | }; |
353 | |
354 | // String based context id |
355 | struct StringBasedCtxKey : public ContextKey { |
356 | SampleContextFrameVector Context; |
357 | |
358 | bool WasLeafInlined; |
359 | StringBasedCtxKey() : ContextKey(CK_StringBased), WasLeafInlined(false){}; |
360 | static bool classof(const ContextKey *K) { |
361 | return K->getKind() == CK_StringBased; |
362 | } |
363 | |
364 | bool isEqual(const ContextKey *K) const override { |
365 | const StringBasedCtxKey *Other = dyn_cast<StringBasedCtxKey>(Val: K); |
366 | return Context == Other->Context; |
367 | } |
368 | |
369 | void genHashCode() override { |
370 | HashCode = hash_value(S: SampleContextFrames(Context)); |
371 | } |
372 | }; |
373 | |
374 | // Address-based context id |
375 | struct AddrBasedCtxKey : public ContextKey { |
376 | SmallVector<uint64_t, 16> Context; |
377 | |
378 | bool WasLeafInlined; |
379 | AddrBasedCtxKey() : ContextKey(CK_AddrBased), WasLeafInlined(false){}; |
380 | static bool classof(const ContextKey *K) { |
381 | return K->getKind() == CK_AddrBased; |
382 | } |
383 | |
384 | bool isEqual(const ContextKey *K) const override { |
385 | const AddrBasedCtxKey *Other = dyn_cast<AddrBasedCtxKey>(Val: K); |
386 | return Context == Other->Context; |
387 | } |
388 | |
389 | void genHashCode() override { |
390 | HashCode = hash_combine_range(first: Context.begin(), last: Context.end()); |
391 | } |
392 | }; |
393 | |
394 | // The counter of branch samples for one function indexed by the branch, |
395 | // which is represented as the source and target offset pair. |
396 | using BranchSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
397 | // The counter of range samples for one function indexed by the range, |
398 | // which is represented as the start and end offset pair. |
399 | using RangeSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
400 | // Wrapper for sample counters including range counter and branch counter |
401 | struct SampleCounter { |
402 | RangeSample RangeCounter; |
403 | BranchSample BranchCounter; |
404 | |
405 | void recordRangeCount(uint64_t Start, uint64_t End, uint64_t Repeat) { |
406 | assert(Start <= End && "Invalid instruction range" ); |
407 | RangeCounter[{Start, End}] += Repeat; |
408 | } |
409 | void recordBranchCount(uint64_t Source, uint64_t Target, uint64_t Repeat) { |
410 | BranchCounter[{Source, Target}] += Repeat; |
411 | } |
412 | }; |
413 | |
414 | // Sample counter with context to support context-sensitive profile |
415 | using ContextSampleCounterMap = |
416 | std::unordered_map<Hashable<ContextKey>, SampleCounter, |
417 | Hashable<ContextKey>::Hash, Hashable<ContextKey>::Equal>; |
418 | |
419 | struct FrameStack { |
420 | SmallVector<uint64_t, 16> Stack; |
421 | ProfiledBinary *Binary; |
422 | FrameStack(ProfiledBinary *B) : Binary(B) {} |
423 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
424 | assert(!Cur->isExternalFrame() && |
425 | "External frame's not expected for context stack." ); |
426 | Stack.push_back(Elt: Cur->Address); |
427 | return true; |
428 | } |
429 | |
430 | void popFrame() { |
431 | if (!Stack.empty()) |
432 | Stack.pop_back(); |
433 | } |
434 | std::shared_ptr<StringBasedCtxKey> getContextKey(); |
435 | }; |
436 | |
437 | struct AddressStack { |
438 | SmallVector<uint64_t, 16> Stack; |
439 | ProfiledBinary *Binary; |
440 | AddressStack(ProfiledBinary *B) : Binary(B) {} |
441 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
442 | assert(!Cur->isExternalFrame() && |
443 | "External frame's not expected for context stack." ); |
444 | Stack.push_back(Elt: Cur->Address); |
445 | return true; |
446 | } |
447 | |
448 | void popFrame() { |
449 | if (!Stack.empty()) |
450 | Stack.pop_back(); |
451 | } |
452 | std::shared_ptr<AddrBasedCtxKey> getContextKey(); |
453 | }; |
454 | |
455 | /* |
456 | As in hybrid sample we have a group of LBRs and the most recent sampling call |
457 | stack, we can walk through those LBRs to infer more call stacks which would be |
458 | used as context for profile. VirtualUnwinder is the class to do the call stack |
459 | unwinding based on LBR state. Two types of unwinding are processd here: |
460 | 1) LBR unwinding and 2) linear range unwinding. |
461 | Specifically, for each LBR entry(can be classified into call, return, regular |
462 | branch), LBR unwinding will replay the operation by pushing, popping or |
463 | switching leaf frame towards the call stack and since the initial call stack |
464 | is most recently sampled, the replay should be in anti-execution order, i.e. for |
465 | the regular case, pop the call stack when LBR is call, push frame on call stack |
466 | when LBR is return. After each LBR processed, it also needs to align with the |
467 | next LBR by going through instructions from previous LBR's target to current |
468 | LBR's source, which is the linear unwinding. As instruction from linear range |
469 | can come from different function by inlining, linear unwinding will do the range |
470 | splitting and record counters by the range with same inline context. Over those |
471 | unwinding process we will record each call stack as context id and LBR/linear |
472 | range as sample counter for further CS profile generation. |
473 | */ |
474 | class VirtualUnwinder { |
475 | public: |
476 | VirtualUnwinder(ContextSampleCounterMap *Counter, ProfiledBinary *B) |
477 | : CtxCounterMap(Counter), Binary(B) {} |
478 | bool unwind(const PerfSample *Sample, uint64_t Repeat); |
479 | std::set<uint64_t> &getUntrackedCallsites() { return UntrackedCallsites; } |
480 | |
481 | uint64_t NumTotalBranches = 0; |
482 | uint64_t NumExtCallBranch = 0; |
483 | uint64_t NumMissingExternalFrame = 0; |
484 | uint64_t NumMismatchedProEpiBranch = 0; |
485 | uint64_t NumMismatchedExtCallBranch = 0; |
486 | uint64_t NumUnpairedExtAddr = 0; |
487 | uint64_t NumPairedExtAddr = 0; |
488 | |
489 | private: |
490 | bool isSourceExternal(UnwindState &State) const { |
491 | return State.getCurrentLBRSource() == ExternalAddr; |
492 | } |
493 | |
494 | bool isTargetExternal(UnwindState &State) const { |
495 | return State.getCurrentLBRTarget() == ExternalAddr; |
496 | } |
497 | |
498 | // Determine whether the return source is from external code by checking if |
499 | // the target's the next inst is a call inst. |
500 | bool isReturnFromExternal(UnwindState &State) const { |
501 | return isSourceExternal(State) && |
502 | (Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) != 0); |
503 | } |
504 | |
505 | // If the source is external address but it's not the `return` case, treat it |
506 | // as a call from external. |
507 | bool isCallFromExternal(UnwindState &State) const { |
508 | return isSourceExternal(State) && |
509 | Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) == 0; |
510 | } |
511 | |
512 | bool isCallState(UnwindState &State) const { |
513 | // The tail call frame is always missing here in stack sample, we will |
514 | // use a specific tail call tracker to infer it. |
515 | if (!isValidState(State)) |
516 | return false; |
517 | |
518 | if (Binary->addressIsCall(Address: State.getCurrentLBRSource())) |
519 | return true; |
520 | |
521 | return isCallFromExternal(State); |
522 | } |
523 | |
524 | bool isReturnState(UnwindState &State) const { |
525 | if (!isValidState(State)) |
526 | return false; |
527 | |
528 | // Simply check addressIsReturn, as ret is always reliable, both for |
529 | // regular call and tail call. |
530 | if (Binary->addressIsReturn(Address: State.getCurrentLBRSource())) |
531 | return true; |
532 | |
533 | return isReturnFromExternal(State); |
534 | } |
535 | |
536 | bool isValidState(UnwindState &State) const { return !State.Invalid; } |
537 | |
538 | void unwindCall(UnwindState &State); |
539 | void unwindLinear(UnwindState &State, uint64_t Repeat); |
540 | void unwindReturn(UnwindState &State); |
541 | void unwindBranch(UnwindState &State); |
542 | |
543 | template <typename T> |
544 | void collectSamplesFromFrame(UnwindState::ProfiledFrame *Cur, T &Stack); |
545 | // Collect each samples on trie node by DFS traversal |
546 | template <typename T> |
547 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur, T &Stack); |
548 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur); |
549 | |
550 | void recordRangeCount(uint64_t Start, uint64_t End, UnwindState &State, |
551 | uint64_t Repeat); |
552 | void recordBranchCount(const LBREntry &Branch, UnwindState &State, |
553 | uint64_t Repeat); |
554 | |
555 | ContextSampleCounterMap *CtxCounterMap; |
556 | // Profiled binary that current frame address belongs to |
557 | ProfiledBinary *Binary; |
558 | // Keep track of all untracked callsites |
559 | std::set<uint64_t> UntrackedCallsites; |
560 | }; |
561 | |
562 | // Read perf trace to parse the events and samples. |
563 | class PerfReaderBase { |
564 | public: |
565 | PerfReaderBase(ProfiledBinary *B, StringRef PerfTrace) |
566 | : Binary(B), PerfTraceFile(PerfTrace) { |
567 | // Initialize the base address to preferred address. |
568 | Binary->setBaseAddress(Binary->getPreferredBaseAddress()); |
569 | }; |
570 | virtual ~PerfReaderBase() = default; |
571 | static std::unique_ptr<PerfReaderBase> |
572 | create(ProfiledBinary *Binary, PerfInputFile &PerfInput, |
573 | std::optional<int32_t> PIDFilter); |
574 | |
575 | // Entry of the reader to parse multiple perf traces |
576 | virtual void parsePerfTraces() = 0; |
577 | const ContextSampleCounterMap &getSampleCounters() const { |
578 | return SampleCounters; |
579 | } |
580 | bool profileIsCS() { return ProfileIsCS; } |
581 | |
582 | protected: |
583 | ProfiledBinary *Binary = nullptr; |
584 | StringRef PerfTraceFile; |
585 | |
586 | ContextSampleCounterMap SampleCounters; |
587 | bool ProfileIsCS = false; |
588 | |
589 | uint64_t NumTotalSample = 0; |
590 | uint64_t NumLeafExternalFrame = 0; |
591 | uint64_t NumLeadingOutgoingLBR = 0; |
592 | }; |
593 | |
594 | // Read perf script to parse the events and samples. |
595 | class PerfScriptReader : public PerfReaderBase { |
596 | public: |
597 | PerfScriptReader(ProfiledBinary *B, StringRef PerfTrace, |
598 | std::optional<int32_t> PID) |
599 | : PerfReaderBase(B, PerfTrace), PIDFilter(PID) {}; |
600 | |
601 | // Entry of the reader to parse multiple perf traces |
602 | void parsePerfTraces() override; |
603 | // Generate perf script from perf data |
604 | static PerfInputFile convertPerfDataToTrace(ProfiledBinary *Binary, |
605 | bool SkipPID, PerfInputFile &File, |
606 | std::optional<int32_t> PIDFilter); |
607 | // Extract perf script type by peaking at the input |
608 | static PerfContent checkPerfScriptType(StringRef FileName); |
609 | |
610 | // Cleanup installers for temporary files created by perf script command. |
611 | // Those files will be automatically removed when running destructor or |
612 | // receiving signals. |
613 | static SmallVector<CleanupInstaller, 2> TempFileCleanups; |
614 | |
615 | protected: |
616 | // The parsed MMap event |
617 | struct MMapEvent { |
618 | int64_t PID = 0; |
619 | uint64_t Address = 0; |
620 | uint64_t Size = 0; |
621 | uint64_t Offset = 0; |
622 | StringRef BinaryPath; |
623 | }; |
624 | |
625 | // Check whether a given line is LBR sample |
626 | static bool isLBRSample(StringRef Line); |
627 | // Check whether a given line is MMAP event |
628 | static bool isMMapEvent(StringRef Line); |
629 | // Parse a single line of a PERF_RECORD_MMAP event looking for a |
630 | // mapping between the binary name and its memory layout. |
631 | static bool (ProfiledBinary *Binary, StringRef Line, |
632 | MMapEvent &MMap); |
633 | // Update base address based on mmap events |
634 | void updateBinaryAddress(const MMapEvent &Event); |
635 | // Parse mmap event and update binary address |
636 | void parseMMapEvent(TraceStream &TraceIt); |
637 | // Parse perf events/samples and do aggregation |
638 | void parseAndAggregateTrace(); |
639 | // Parse either an MMAP event or a perf sample |
640 | void parseEventOrSample(TraceStream &TraceIt); |
641 | // Warn if the relevant mmap event is missing. |
642 | void warnIfMissingMMap(); |
643 | // Emit accumulate warnings. |
644 | void warnTruncatedStack(); |
645 | // Warn if range is invalid. |
646 | void warnInvalidRange(); |
647 | // Extract call stack from the perf trace lines |
648 | bool (TraceStream &TraceIt, |
649 | SmallVectorImpl<uint64_t> &CallStack); |
650 | // Extract LBR stack from one perf trace line |
651 | bool (TraceStream &TraceIt, |
652 | SmallVectorImpl<LBREntry> &LBRStack); |
653 | uint64_t parseAggregatedCount(TraceStream &TraceIt); |
654 | // Parse one sample from multiple perf lines, override this for different |
655 | // sample type |
656 | void parseSample(TraceStream &TraceIt); |
657 | // An aggregated count is given to indicate how many times the sample is |
658 | // repeated. |
659 | virtual void parseSample(TraceStream &TraceIt, uint64_t Count){}; |
660 | void computeCounterFromLBR(const PerfSample *Sample, uint64_t Repeat); |
661 | // Post process the profile after trace aggregation, we will do simple range |
662 | // overlap computation for AutoFDO, or unwind for CSSPGO(hybrid sample). |
663 | virtual void generateUnsymbolizedProfile(); |
664 | void writeUnsymbolizedProfile(StringRef Filename); |
665 | void writeUnsymbolizedProfile(raw_fd_ostream &OS); |
666 | |
667 | // Samples with the repeating time generated by the perf reader |
668 | AggregatedCounter AggregatedSamples; |
669 | // Keep track of all invalid return addresses |
670 | std::set<uint64_t> InvalidReturnAddresses; |
671 | // PID for the process of interest |
672 | std::optional<int32_t> PIDFilter; |
673 | }; |
674 | |
675 | /* |
676 | The reader of LBR only perf script. |
677 | A typical LBR sample is like: |
678 | 40062f 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
679 | ... 0x4005c8/0x4005dc/P/-/-/0 |
680 | */ |
681 | class LBRPerfReader : public PerfScriptReader { |
682 | public: |
683 | LBRPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
684 | std::optional<int32_t> PID) |
685 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
686 | // Parse the LBR only sample. |
687 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
688 | }; |
689 | |
690 | /* |
691 | Hybrid perf script includes a group of hybrid samples(LBRs + call stack), |
692 | which is used to generate CS profile. An example of hybrid sample: |
693 | 4005dc # call stack leaf |
694 | 400634 |
695 | 400684 # call stack root |
696 | 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
697 | ... 0x4005c8/0x4005dc/P/-/-/0 # LBR Entries |
698 | */ |
699 | class HybridPerfReader : public PerfScriptReader { |
700 | public: |
701 | HybridPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
702 | std::optional<int32_t> PID) |
703 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
704 | // Parse the hybrid sample including the call and LBR line |
705 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
706 | void generateUnsymbolizedProfile() override; |
707 | |
708 | private: |
709 | // Unwind the hybrid samples after aggregration |
710 | void unwindSamples(); |
711 | }; |
712 | |
713 | /* |
714 | Format of unsymbolized profile: |
715 | |
716 | [frame1 @ frame2 @ ...] # If it's a CS profile |
717 | number of entries in RangeCounter |
718 | from_1-to_1:count_1 |
719 | from_2-to_2:count_2 |
720 | ...... |
721 | from_n-to_n:count_n |
722 | number of entries in BranchCounter |
723 | src_1->dst_1:count_1 |
724 | src_2->dst_2:count_2 |
725 | ...... |
726 | src_n->dst_n:count_n |
727 | [frame1 @ frame2 @ ...] # Next context |
728 | ...... |
729 | |
730 | Note that non-CS profile doesn't have the empty `[]` context. |
731 | */ |
732 | class UnsymbolizedProfileReader : public PerfReaderBase { |
733 | public: |
734 | UnsymbolizedProfileReader(ProfiledBinary *Binary, StringRef PerfTrace) |
735 | : PerfReaderBase(Binary, PerfTrace){}; |
736 | void parsePerfTraces() override; |
737 | |
738 | private: |
739 | void readSampleCounters(TraceStream &TraceIt, SampleCounter &SCounters); |
740 | void readUnsymbolizedProfile(StringRef Filename); |
741 | |
742 | std::unordered_set<std::string> ContextStrSet; |
743 | }; |
744 | |
745 | } // end namespace sampleprof |
746 | } // end namespace llvm |
747 | |
748 | #endif |
749 | |