| 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(first1: CallStack.begin(), last1: CallStack.end(), first2: 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 { HashCode = hash_combine_range(R&: Context); } |
| 390 | }; |
| 391 | |
| 392 | // The counter of branch samples for one function indexed by the branch, |
| 393 | // which is represented as the source and target offset pair. |
| 394 | using BranchSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
| 395 | // The counter of range samples for one function indexed by the range, |
| 396 | // which is represented as the start and end offset pair. |
| 397 | using RangeSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
| 398 | // Wrapper for sample counters including range counter and branch counter |
| 399 | struct SampleCounter { |
| 400 | RangeSample RangeCounter; |
| 401 | BranchSample BranchCounter; |
| 402 | |
| 403 | void recordRangeCount(uint64_t Start, uint64_t End, uint64_t Repeat) { |
| 404 | assert(Start <= End && "Invalid instruction range" ); |
| 405 | RangeCounter[{Start, End}] += Repeat; |
| 406 | } |
| 407 | void recordBranchCount(uint64_t Source, uint64_t Target, uint64_t Repeat) { |
| 408 | BranchCounter[{Source, Target}] += Repeat; |
| 409 | } |
| 410 | }; |
| 411 | |
| 412 | // Sample counter with context to support context-sensitive profile |
| 413 | using ContextSampleCounterMap = |
| 414 | std::unordered_map<Hashable<ContextKey>, SampleCounter, |
| 415 | Hashable<ContextKey>::Hash, Hashable<ContextKey>::Equal>; |
| 416 | |
| 417 | struct FrameStack { |
| 418 | SmallVector<uint64_t, 16> Stack; |
| 419 | ProfiledBinary *Binary; |
| 420 | FrameStack(ProfiledBinary *B) : Binary(B) {} |
| 421 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
| 422 | assert(!Cur->isExternalFrame() && |
| 423 | "External frame's not expected for context stack." ); |
| 424 | Stack.push_back(Elt: Cur->Address); |
| 425 | return true; |
| 426 | } |
| 427 | |
| 428 | void popFrame() { |
| 429 | if (!Stack.empty()) |
| 430 | Stack.pop_back(); |
| 431 | } |
| 432 | std::shared_ptr<StringBasedCtxKey> getContextKey(); |
| 433 | }; |
| 434 | |
| 435 | struct AddressStack { |
| 436 | SmallVector<uint64_t, 16> Stack; |
| 437 | ProfiledBinary *Binary; |
| 438 | AddressStack(ProfiledBinary *B) : Binary(B) {} |
| 439 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
| 440 | assert(!Cur->isExternalFrame() && |
| 441 | "External frame's not expected for context stack." ); |
| 442 | Stack.push_back(Elt: Cur->Address); |
| 443 | return true; |
| 444 | } |
| 445 | |
| 446 | void popFrame() { |
| 447 | if (!Stack.empty()) |
| 448 | Stack.pop_back(); |
| 449 | } |
| 450 | std::shared_ptr<AddrBasedCtxKey> getContextKey(); |
| 451 | }; |
| 452 | |
| 453 | /* |
| 454 | As in hybrid sample we have a group of LBRs and the most recent sampling call |
| 455 | stack, we can walk through those LBRs to infer more call stacks which would be |
| 456 | used as context for profile. VirtualUnwinder is the class to do the call stack |
| 457 | unwinding based on LBR state. Two types of unwinding are processd here: |
| 458 | 1) LBR unwinding and 2) linear range unwinding. |
| 459 | Specifically, for each LBR entry(can be classified into call, return, regular |
| 460 | branch), LBR unwinding will replay the operation by pushing, popping or |
| 461 | switching leaf frame towards the call stack and since the initial call stack |
| 462 | is most recently sampled, the replay should be in anti-execution order, i.e. for |
| 463 | the regular case, pop the call stack when LBR is call, push frame on call stack |
| 464 | when LBR is return. After each LBR processed, it also needs to align with the |
| 465 | next LBR by going through instructions from previous LBR's target to current |
| 466 | LBR's source, which is the linear unwinding. As instruction from linear range |
| 467 | can come from different function by inlining, linear unwinding will do the range |
| 468 | splitting and record counters by the range with same inline context. Over those |
| 469 | unwinding process we will record each call stack as context id and LBR/linear |
| 470 | range as sample counter for further CS profile generation. |
| 471 | */ |
| 472 | class VirtualUnwinder { |
| 473 | public: |
| 474 | VirtualUnwinder(ContextSampleCounterMap *Counter, ProfiledBinary *B) |
| 475 | : CtxCounterMap(Counter), Binary(B) {} |
| 476 | bool unwind(const PerfSample *Sample, uint64_t Repeat); |
| 477 | std::set<uint64_t> &getUntrackedCallsites() { return UntrackedCallsites; } |
| 478 | |
| 479 | uint64_t NumTotalBranches = 0; |
| 480 | uint64_t NumExtCallBranch = 0; |
| 481 | uint64_t NumMissingExternalFrame = 0; |
| 482 | uint64_t NumMismatchedProEpiBranch = 0; |
| 483 | uint64_t NumMismatchedExtCallBranch = 0; |
| 484 | uint64_t NumUnpairedExtAddr = 0; |
| 485 | uint64_t NumPairedExtAddr = 0; |
| 486 | |
| 487 | private: |
| 488 | bool isSourceExternal(UnwindState &State) const { |
| 489 | return State.getCurrentLBRSource() == ExternalAddr; |
| 490 | } |
| 491 | |
| 492 | bool isTargetExternal(UnwindState &State) const { |
| 493 | return State.getCurrentLBRTarget() == ExternalAddr; |
| 494 | } |
| 495 | |
| 496 | // Determine whether the return source is from external code by checking if |
| 497 | // the target's the next inst is a call inst. |
| 498 | bool isReturnFromExternal(UnwindState &State) const { |
| 499 | return isSourceExternal(State) && |
| 500 | (Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) != 0); |
| 501 | } |
| 502 | |
| 503 | // If the source is external address but it's not the `return` case, treat it |
| 504 | // as a call from external. |
| 505 | bool isCallFromExternal(UnwindState &State) const { |
| 506 | return isSourceExternal(State) && |
| 507 | Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) == 0; |
| 508 | } |
| 509 | |
| 510 | bool isCallState(UnwindState &State) const { |
| 511 | // The tail call frame is always missing here in stack sample, we will |
| 512 | // use a specific tail call tracker to infer it. |
| 513 | if (!isValidState(State)) |
| 514 | return false; |
| 515 | |
| 516 | if (Binary->addressIsCall(Address: State.getCurrentLBRSource())) |
| 517 | return true; |
| 518 | |
| 519 | return isCallFromExternal(State); |
| 520 | } |
| 521 | |
| 522 | bool isReturnState(UnwindState &State) const { |
| 523 | if (!isValidState(State)) |
| 524 | return false; |
| 525 | |
| 526 | // Simply check addressIsReturn, as ret is always reliable, both for |
| 527 | // regular call and tail call. |
| 528 | if (Binary->addressIsReturn(Address: State.getCurrentLBRSource())) |
| 529 | return true; |
| 530 | |
| 531 | return isReturnFromExternal(State); |
| 532 | } |
| 533 | |
| 534 | bool isValidState(UnwindState &State) const { return !State.Invalid; } |
| 535 | |
| 536 | void unwindCall(UnwindState &State); |
| 537 | void unwindLinear(UnwindState &State, uint64_t Repeat); |
| 538 | void unwindReturn(UnwindState &State); |
| 539 | void unwindBranch(UnwindState &State); |
| 540 | |
| 541 | template <typename T> |
| 542 | void collectSamplesFromFrame(UnwindState::ProfiledFrame *Cur, T &Stack); |
| 543 | // Collect each samples on trie node by DFS traversal |
| 544 | template <typename T> |
| 545 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur, T &Stack); |
| 546 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur); |
| 547 | |
| 548 | void recordRangeCount(uint64_t Start, uint64_t End, UnwindState &State, |
| 549 | uint64_t Repeat); |
| 550 | void recordBranchCount(const LBREntry &Branch, UnwindState &State, |
| 551 | uint64_t Repeat); |
| 552 | |
| 553 | ContextSampleCounterMap *CtxCounterMap; |
| 554 | // Profiled binary that current frame address belongs to |
| 555 | ProfiledBinary *Binary; |
| 556 | // Keep track of all untracked callsites |
| 557 | std::set<uint64_t> UntrackedCallsites; |
| 558 | }; |
| 559 | |
| 560 | // Read perf trace to parse the events and samples. |
| 561 | class PerfReaderBase { |
| 562 | public: |
| 563 | PerfReaderBase(ProfiledBinary *B, StringRef PerfTrace) |
| 564 | : Binary(B), PerfTraceFile(PerfTrace) { |
| 565 | // Initialize the base address to preferred address. |
| 566 | Binary->setBaseAddress(Binary->getPreferredBaseAddress()); |
| 567 | }; |
| 568 | virtual ~PerfReaderBase() = default; |
| 569 | static std::unique_ptr<PerfReaderBase> |
| 570 | create(ProfiledBinary *Binary, PerfInputFile &PerfInput, |
| 571 | std::optional<int32_t> PIDFilter); |
| 572 | |
| 573 | // Entry of the reader to parse multiple perf traces |
| 574 | virtual void parsePerfTraces() = 0; |
| 575 | const ContextSampleCounterMap &getSampleCounters() const { |
| 576 | return SampleCounters; |
| 577 | } |
| 578 | bool profileIsCS() { return ProfileIsCS; } |
| 579 | |
| 580 | protected: |
| 581 | ProfiledBinary *Binary = nullptr; |
| 582 | StringRef PerfTraceFile; |
| 583 | |
| 584 | ContextSampleCounterMap SampleCounters; |
| 585 | bool ProfileIsCS = false; |
| 586 | |
| 587 | uint64_t NumTotalSample = 0; |
| 588 | uint64_t NumLeafExternalFrame = 0; |
| 589 | uint64_t NumLeadingOutgoingLBR = 0; |
| 590 | }; |
| 591 | |
| 592 | // Read perf script to parse the events and samples. |
| 593 | class PerfScriptReader : public PerfReaderBase { |
| 594 | public: |
| 595 | PerfScriptReader(ProfiledBinary *B, StringRef PerfTrace, |
| 596 | std::optional<int32_t> PID) |
| 597 | : PerfReaderBase(B, PerfTrace), PIDFilter(PID) {}; |
| 598 | |
| 599 | // Entry of the reader to parse multiple perf traces |
| 600 | void parsePerfTraces() override; |
| 601 | // Generate perf script from perf data |
| 602 | static PerfInputFile convertPerfDataToTrace(ProfiledBinary *Binary, |
| 603 | bool SkipPID, PerfInputFile &File, |
| 604 | std::optional<int32_t> PIDFilter); |
| 605 | // Extract perf script type by peaking at the input |
| 606 | static PerfContent checkPerfScriptType(StringRef FileName); |
| 607 | |
| 608 | // Cleanup installers for temporary files created by perf script command. |
| 609 | // Those files will be automatically removed when running destructor or |
| 610 | // receiving signals. |
| 611 | static SmallVector<CleanupInstaller, 2> TempFileCleanups; |
| 612 | |
| 613 | protected: |
| 614 | // The parsed MMap event |
| 615 | struct MMapEvent { |
| 616 | int64_t PID = 0; |
| 617 | uint64_t Address = 0; |
| 618 | uint64_t Size = 0; |
| 619 | uint64_t Offset = 0; |
| 620 | StringRef BinaryPath; |
| 621 | }; |
| 622 | |
| 623 | // Check whether a given line is LBR sample |
| 624 | static bool isLBRSample(StringRef Line); |
| 625 | // Check whether a given line is MMAP event |
| 626 | static bool isMMapEvent(StringRef Line); |
| 627 | // Parse a single line of a PERF_RECORD_MMAP event looking for a |
| 628 | // mapping between the binary name and its memory layout. |
| 629 | static bool (ProfiledBinary *Binary, StringRef Line, |
| 630 | MMapEvent &MMap); |
| 631 | // Update base address based on mmap events |
| 632 | void updateBinaryAddress(const MMapEvent &Event); |
| 633 | // Parse mmap event and update binary address |
| 634 | void parseMMapEvent(TraceStream &TraceIt); |
| 635 | // Parse perf events/samples and do aggregation |
| 636 | void parseAndAggregateTrace(); |
| 637 | // Parse either an MMAP event or a perf sample |
| 638 | void parseEventOrSample(TraceStream &TraceIt); |
| 639 | // Warn if the relevant mmap event is missing. |
| 640 | void warnIfMissingMMap(); |
| 641 | // Emit accumulate warnings. |
| 642 | void warnTruncatedStack(); |
| 643 | // Warn if range is invalid. |
| 644 | void warnInvalidRange(); |
| 645 | // Extract call stack from the perf trace lines |
| 646 | bool (TraceStream &TraceIt, |
| 647 | SmallVectorImpl<uint64_t> &CallStack); |
| 648 | // Extract LBR stack from one perf trace line |
| 649 | bool (TraceStream &TraceIt, |
| 650 | SmallVectorImpl<LBREntry> &LBRStack); |
| 651 | uint64_t parseAggregatedCount(TraceStream &TraceIt); |
| 652 | // Parse one sample from multiple perf lines, override this for different |
| 653 | // sample type |
| 654 | void parseSample(TraceStream &TraceIt); |
| 655 | // An aggregated count is given to indicate how many times the sample is |
| 656 | // repeated. |
| 657 | virtual void parseSample(TraceStream &TraceIt, uint64_t Count){}; |
| 658 | void computeCounterFromLBR(const PerfSample *Sample, uint64_t Repeat); |
| 659 | // Post process the profile after trace aggregation, we will do simple range |
| 660 | // overlap computation for AutoFDO, or unwind for CSSPGO(hybrid sample). |
| 661 | virtual void generateUnsymbolizedProfile(); |
| 662 | void writeUnsymbolizedProfile(StringRef Filename); |
| 663 | void writeUnsymbolizedProfile(raw_fd_ostream &OS); |
| 664 | |
| 665 | // Samples with the repeating time generated by the perf reader |
| 666 | AggregatedCounter AggregatedSamples; |
| 667 | // Keep track of all invalid return addresses |
| 668 | std::set<uint64_t> InvalidReturnAddresses; |
| 669 | // PID for the process of interest |
| 670 | std::optional<int32_t> PIDFilter; |
| 671 | }; |
| 672 | |
| 673 | /* |
| 674 | The reader of LBR only perf script. |
| 675 | A typical LBR sample is like: |
| 676 | 40062f 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
| 677 | ... 0x4005c8/0x4005dc/P/-/-/0 |
| 678 | */ |
| 679 | class LBRPerfReader : public PerfScriptReader { |
| 680 | public: |
| 681 | LBRPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
| 682 | std::optional<int32_t> PID) |
| 683 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
| 684 | // Parse the LBR only sample. |
| 685 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
| 686 | }; |
| 687 | |
| 688 | /* |
| 689 | Hybrid perf script includes a group of hybrid samples(LBRs + call stack), |
| 690 | which is used to generate CS profile. An example of hybrid sample: |
| 691 | 4005dc # call stack leaf |
| 692 | 400634 |
| 693 | 400684 # call stack root |
| 694 | 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
| 695 | ... 0x4005c8/0x4005dc/P/-/-/0 # LBR Entries |
| 696 | */ |
| 697 | class HybridPerfReader : public PerfScriptReader { |
| 698 | public: |
| 699 | HybridPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
| 700 | std::optional<int32_t> PID) |
| 701 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
| 702 | // Parse the hybrid sample including the call and LBR line |
| 703 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
| 704 | void generateUnsymbolizedProfile() override; |
| 705 | |
| 706 | private: |
| 707 | // Unwind the hybrid samples after aggregration |
| 708 | void unwindSamples(); |
| 709 | }; |
| 710 | |
| 711 | /* |
| 712 | Format of unsymbolized profile: |
| 713 | |
| 714 | [frame1 @ frame2 @ ...] # If it's a CS profile |
| 715 | number of entries in RangeCounter |
| 716 | from_1-to_1:count_1 |
| 717 | from_2-to_2:count_2 |
| 718 | ...... |
| 719 | from_n-to_n:count_n |
| 720 | number of entries in BranchCounter |
| 721 | src_1->dst_1:count_1 |
| 722 | src_2->dst_2:count_2 |
| 723 | ...... |
| 724 | src_n->dst_n:count_n |
| 725 | [frame1 @ frame2 @ ...] # Next context |
| 726 | ...... |
| 727 | |
| 728 | Note that non-CS profile doesn't have the empty `[]` context. |
| 729 | */ |
| 730 | class UnsymbolizedProfileReader : public PerfReaderBase { |
| 731 | public: |
| 732 | UnsymbolizedProfileReader(ProfiledBinary *Binary, StringRef PerfTrace) |
| 733 | : PerfReaderBase(Binary, PerfTrace){}; |
| 734 | void parsePerfTraces() override; |
| 735 | |
| 736 | private: |
| 737 | void readSampleCounters(TraceStream &TraceIt, SampleCounter &SCounters); |
| 738 | void readUnsymbolizedProfile(StringRef Filename); |
| 739 | |
| 740 | std::unordered_set<std::string> ContextStrSet; |
| 741 | }; |
| 742 | |
| 743 | } // end namespace sampleprof |
| 744 | } // end namespace llvm |
| 745 | |
| 746 | #endif |
| 747 | |