| 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/Error.h" |
| 16 | #include "llvm/Support/Regex.h" |
| 17 | #include <cstdint> |
| 18 | #include <fstream> |
| 19 | #include <map> |
| 20 | |
| 21 | namespace llvm { |
| 22 | |
| 23 | class CleanupInstaller; |
| 24 | |
| 25 | namespace sampleprof { |
| 26 | |
| 27 | // Stream based trace line iterator |
| 28 | class TraceStream { |
| 29 | std::string CurrentLine; |
| 30 | std::ifstream Fin; |
| 31 | bool IsAtEoF = false; |
| 32 | uint64_t LineNumber = 0; |
| 33 | |
| 34 | public: |
| 35 | TraceStream(StringRef Filename) : Fin(Filename.str()) { |
| 36 | if (!Fin.good()) |
| 37 | exitWithError(Message: "Error read input perf script file" , Whence: Filename); |
| 38 | advance(); |
| 39 | } |
| 40 | |
| 41 | StringRef getCurrentLine() { |
| 42 | assert(!IsAtEoF && "Line iterator reaches the End-of-File!" ); |
| 43 | return CurrentLine; |
| 44 | } |
| 45 | |
| 46 | uint64_t getLineNumber() { return LineNumber; } |
| 47 | |
| 48 | bool isAtEoF() { return IsAtEoF; } |
| 49 | |
| 50 | // Read the next line |
| 51 | void advance() { |
| 52 | if (!std::getline(is&: Fin, str&: CurrentLine)) { |
| 53 | IsAtEoF = true; |
| 54 | return; |
| 55 | } |
| 56 | LineNumber++; |
| 57 | } |
| 58 | }; |
| 59 | |
| 60 | // The type of input format. |
| 61 | enum PerfFormat { |
| 62 | UnknownFormat = 0, |
| 63 | PerfData = 1, // Raw linux perf.data. |
| 64 | PerfScript = 2, // Perf script create by `perf script` command. |
| 65 | UnsymbolizedProfile = 3, // Unsymbolized profile generated by llvm-profgen. |
| 66 | |
| 67 | }; |
| 68 | |
| 69 | // The type of perfscript content. |
| 70 | enum PerfContent { |
| 71 | UnknownContent = 0, |
| 72 | LBR = 1, // Only LBR sample. |
| 73 | LBRStack = 2, // Hybrid sample including call stack and LBR stack. |
| 74 | }; |
| 75 | |
| 76 | struct PerfInputFile { |
| 77 | std::string InputFile; |
| 78 | PerfFormat Format = PerfFormat::UnknownFormat; |
| 79 | PerfContent Content = PerfContent::UnknownContent; |
| 80 | }; |
| 81 | |
| 82 | // The parsed LBR sample entry. |
| 83 | struct LBREntry { |
| 84 | uint64_t Source = 0; |
| 85 | uint64_t Target = 0; |
| 86 | LBREntry(uint64_t S, uint64_t T) : Source(S), Target(T) {} |
| 87 | |
| 88 | #ifndef NDEBUG |
| 89 | void print() const { |
| 90 | dbgs() << "from " << format("%#010x" , Source) << " to " |
| 91 | << format("%#010x" , Target); |
| 92 | } |
| 93 | #endif |
| 94 | }; |
| 95 | |
| 96 | #ifndef NDEBUG |
| 97 | static inline void printLBRStack(const SmallVectorImpl<LBREntry> &LBRStack) { |
| 98 | for (size_t I = 0; I < LBRStack.size(); I++) { |
| 99 | dbgs() << "[" << I << "] " ; |
| 100 | LBRStack[I].print(); |
| 101 | dbgs() << "\n" ; |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | static inline void printCallStack(const SmallVectorImpl<uint64_t> &CallStack) { |
| 106 | for (size_t I = 0; I < CallStack.size(); I++) { |
| 107 | dbgs() << "[" << I << "] " << format("%#010x" , CallStack[I]) << "\n" ; |
| 108 | } |
| 109 | } |
| 110 | #endif |
| 111 | |
| 112 | // Hash interface for generic data of type T |
| 113 | // Data should implement a \fn getHashCode and a \fn isEqual |
| 114 | // Currently getHashCode is non-virtual to avoid the overhead of calling vtable, |
| 115 | // i.e we explicitly calculate hash of derived class, assign to base class's |
| 116 | // HashCode. This also provides the flexibility for calculating the hash code |
| 117 | // incrementally(like rolling hash) during frame stack unwinding since unwinding |
| 118 | // only changes the leaf of frame stack. \fn isEqual is a virtual function, |
| 119 | // which will have perf overhead. In the future, if we redesign a better hash |
| 120 | // function, then we can just skip this or switch to non-virtual function(like |
| 121 | // just ignore comparison if hash conflicts probabilities is low) |
| 122 | template <class T> class Hashable { |
| 123 | public: |
| 124 | std::shared_ptr<T> Data; |
| 125 | Hashable(const std::shared_ptr<T> &D) : Data(D) {} |
| 126 | |
| 127 | // Hash code generation |
| 128 | struct Hash { |
| 129 | uint64_t operator()(const Hashable<T> &Key) const { |
| 130 | // Don't make it virtual for getHashCode |
| 131 | uint64_t Hash = Key.Data->getHashCode(); |
| 132 | assert(Hash && "Should generate HashCode for it!" ); |
| 133 | return Hash; |
| 134 | } |
| 135 | }; |
| 136 | |
| 137 | // Hash equal |
| 138 | struct Equal { |
| 139 | bool operator()(const Hashable<T> &LHS, const Hashable<T> &RHS) const { |
| 140 | // Precisely compare the data, vtable will have overhead. |
| 141 | return LHS.Data->isEqual(RHS.Data.get()); |
| 142 | } |
| 143 | }; |
| 144 | |
| 145 | T *getPtr() const { return Data.get(); } |
| 146 | }; |
| 147 | |
| 148 | struct PerfSample { |
| 149 | // LBR stack recorded in FIFO order. |
| 150 | SmallVector<LBREntry, 16> LBRStack; |
| 151 | // Call stack recorded in FILO(leaf to root) order, it's used for CS-profile |
| 152 | // generation |
| 153 | SmallVector<uint64_t, 16> CallStack; |
| 154 | |
| 155 | virtual ~PerfSample() = default; |
| 156 | uint64_t getHashCode() const { |
| 157 | // Use simple DJB2 hash |
| 158 | auto HashCombine = [](uint64_t H, uint64_t V) { |
| 159 | return ((H << 5) + H) + V; |
| 160 | }; |
| 161 | uint64_t Hash = 5381; |
| 162 | for (const auto &Value : CallStack) { |
| 163 | Hash = HashCombine(Hash, Value); |
| 164 | } |
| 165 | for (const auto &Entry : LBRStack) { |
| 166 | Hash = HashCombine(Hash, Entry.Source); |
| 167 | Hash = HashCombine(Hash, Entry.Target); |
| 168 | } |
| 169 | return Hash; |
| 170 | } |
| 171 | |
| 172 | bool isEqual(const PerfSample *Other) const { |
| 173 | const SmallVector<uint64_t, 16> &OtherCallStack = Other->CallStack; |
| 174 | const SmallVector<LBREntry, 16> &OtherLBRStack = Other->LBRStack; |
| 175 | |
| 176 | if (CallStack.size() != OtherCallStack.size() || |
| 177 | LBRStack.size() != OtherLBRStack.size()) |
| 178 | return false; |
| 179 | |
| 180 | if (!std::equal(first1: CallStack.begin(), last1: CallStack.end(), first2: OtherCallStack.begin())) |
| 181 | return false; |
| 182 | |
| 183 | for (size_t I = 0; I < OtherLBRStack.size(); I++) { |
| 184 | if (LBRStack[I].Source != OtherLBRStack[I].Source || |
| 185 | LBRStack[I].Target != OtherLBRStack[I].Target) |
| 186 | return false; |
| 187 | } |
| 188 | return true; |
| 189 | } |
| 190 | |
| 191 | #ifndef NDEBUG |
| 192 | uint64_t Linenum = 0; |
| 193 | |
| 194 | void print() const { |
| 195 | dbgs() << "Line " << Linenum << "\n" ; |
| 196 | dbgs() << "LBR stack\n" ; |
| 197 | printLBRStack(LBRStack); |
| 198 | dbgs() << "Call stack\n" ; |
| 199 | printCallStack(CallStack); |
| 200 | } |
| 201 | #endif |
| 202 | }; |
| 203 | // After parsing the sample, we record the samples by aggregating them |
| 204 | // into this counter. The key stores the sample data and the value is |
| 205 | // the sample repeat times. |
| 206 | using AggregatedCounter = |
| 207 | std::unordered_map<Hashable<PerfSample>, uint64_t, |
| 208 | Hashable<PerfSample>::Hash, Hashable<PerfSample>::Equal>; |
| 209 | |
| 210 | using SampleVector = SmallVector<std::tuple<uint64_t, uint64_t, uint64_t>, 16>; |
| 211 | |
| 212 | inline bool isValidFallThroughRange(uint64_t Start, uint64_t End, |
| 213 | ProfiledBinary *Binary) { |
| 214 | // Start bigger than End is considered invalid. |
| 215 | // LBR ranges cross the unconditional jmp are also assumed invalid. |
| 216 | // It's found that perf data may contain duplicate LBR entries that could form |
| 217 | // a range that does not reflect real execution flow on some Intel targets, |
| 218 | // e.g. Skylake. Such ranges are ususally very long. Exclude them since there |
| 219 | // cannot be a linear execution range that spans over unconditional jmp. |
| 220 | return Start <= End && !Binary->rangeCrossUncondBranch(Start, End); |
| 221 | } |
| 222 | |
| 223 | // The state for the unwinder, it doesn't hold the data but only keep the |
| 224 | // pointer/index of the data, While unwinding, the CallStack is changed |
| 225 | // dynamicially and will be recorded as the context of the sample |
| 226 | struct UnwindState { |
| 227 | // Profiled binary that current frame address belongs to |
| 228 | const ProfiledBinary *Binary; |
| 229 | // Call stack trie node |
| 230 | struct ProfiledFrame { |
| 231 | const uint64_t Address = DummyRoot; |
| 232 | ProfiledFrame *Parent; |
| 233 | SampleVector RangeSamples; |
| 234 | SampleVector BranchSamples; |
| 235 | std::unordered_map<uint64_t, std::unique_ptr<ProfiledFrame>> Children; |
| 236 | |
| 237 | ProfiledFrame(uint64_t Addr = 0, ProfiledFrame *P = nullptr) |
| 238 | : Address(Addr), Parent(P) {} |
| 239 | ProfiledFrame *getOrCreateChildFrame(uint64_t Address) { |
| 240 | assert(Address && "Address can't be zero!" ); |
| 241 | auto Ret = Children.emplace( |
| 242 | args&: Address, args: std::make_unique<ProfiledFrame>(args&: Address, args: this)); |
| 243 | return Ret.first->second.get(); |
| 244 | } |
| 245 | void recordRangeCount(uint64_t Start, uint64_t End, uint64_t Count) { |
| 246 | RangeSamples.emplace_back(Args: std::make_tuple(args&: Start, args&: End, args&: Count)); |
| 247 | } |
| 248 | void recordBranchCount(uint64_t Source, uint64_t Target, uint64_t Count) { |
| 249 | BranchSamples.emplace_back(Args: std::make_tuple(args&: Source, args&: Target, args&: Count)); |
| 250 | } |
| 251 | bool isDummyRoot() { return Address == DummyRoot; } |
| 252 | bool isExternalFrame() { return Address == ExternalAddr; } |
| 253 | bool isLeafFrame() { return Children.empty(); } |
| 254 | }; |
| 255 | |
| 256 | ProfiledFrame DummyTrieRoot; |
| 257 | ProfiledFrame *CurrentLeafFrame; |
| 258 | // Used to fall through the LBR stack |
| 259 | uint32_t LBRIndex = 0; |
| 260 | // Reference to PerfSample.LBRStack |
| 261 | const SmallVector<LBREntry, 16> &LBRStack; |
| 262 | // Used to iterate the address range |
| 263 | InstructionPointer InstPtr; |
| 264 | // Indicate whether unwinding is currently in a bad state which requires to |
| 265 | // skip all subsequent unwinding. |
| 266 | bool Invalid = false; |
| 267 | UnwindState(const PerfSample *Sample, const ProfiledBinary *Binary) |
| 268 | : Binary(Binary), LBRStack(Sample->LBRStack), |
| 269 | InstPtr(Binary, Sample->CallStack.front()) { |
| 270 | initFrameTrie(CallStack: Sample->CallStack); |
| 271 | } |
| 272 | |
| 273 | bool validateInitialState() { |
| 274 | uint64_t LBRLeaf = LBRStack[LBRIndex].Target; |
| 275 | uint64_t LeafAddr = CurrentLeafFrame->Address; |
| 276 | assert((LBRLeaf != ExternalAddr || LBRLeaf == LeafAddr) && |
| 277 | "External leading LBR should match the leaf frame." ); |
| 278 | |
| 279 | // When we take a stack sample, ideally the sampling distance between the |
| 280 | // leaf IP of stack and the last LBR target shouldn't be very large. |
| 281 | // Use a heuristic size (0x100) to filter out broken records. |
| 282 | if (LeafAddr < LBRLeaf || LeafAddr - LBRLeaf >= 0x100) { |
| 283 | WithColor::warning() << "Bogus trace: stack tip = " |
| 284 | << format(Fmt: "%#010x" , Vals: LeafAddr) |
| 285 | << ", LBR tip = " << format(Fmt: "%#010x\n" , Vals: LBRLeaf); |
| 286 | return false; |
| 287 | } |
| 288 | return true; |
| 289 | } |
| 290 | |
| 291 | void checkStateConsistency() { |
| 292 | assert(InstPtr.Address == CurrentLeafFrame->Address && |
| 293 | "IP should align with context leaf" ); |
| 294 | } |
| 295 | |
| 296 | void setInvalid() { Invalid = true; } |
| 297 | bool hasNextLBR() const { return LBRIndex < LBRStack.size(); } |
| 298 | uint64_t getCurrentLBRSource() const { return LBRStack[LBRIndex].Source; } |
| 299 | uint64_t getCurrentLBRTarget() const { return LBRStack[LBRIndex].Target; } |
| 300 | const LBREntry &getCurrentLBR() const { return LBRStack[LBRIndex]; } |
| 301 | bool IsLastLBR() const { return LBRIndex == 0; } |
| 302 | bool getLBRStackSize() const { return LBRStack.size(); } |
| 303 | void advanceLBR() { LBRIndex++; } |
| 304 | ProfiledFrame *getParentFrame() { return CurrentLeafFrame->Parent; } |
| 305 | |
| 306 | void pushFrame(uint64_t Address) { |
| 307 | CurrentLeafFrame = CurrentLeafFrame->getOrCreateChildFrame(Address); |
| 308 | } |
| 309 | |
| 310 | void switchToFrame(uint64_t Address) { |
| 311 | if (CurrentLeafFrame->Address == Address) |
| 312 | return; |
| 313 | CurrentLeafFrame = CurrentLeafFrame->Parent->getOrCreateChildFrame(Address); |
| 314 | } |
| 315 | |
| 316 | void popFrame() { CurrentLeafFrame = CurrentLeafFrame->Parent; } |
| 317 | |
| 318 | void clearCallStack() { CurrentLeafFrame = &DummyTrieRoot; } |
| 319 | |
| 320 | void initFrameTrie(const SmallVectorImpl<uint64_t> &CallStack) { |
| 321 | ProfiledFrame *Cur = &DummyTrieRoot; |
| 322 | for (auto Address : reverse(C: CallStack)) { |
| 323 | Cur = Cur->getOrCreateChildFrame(Address); |
| 324 | } |
| 325 | CurrentLeafFrame = Cur; |
| 326 | } |
| 327 | |
| 328 | ProfiledFrame *getDummyRootPtr() { return &DummyTrieRoot; } |
| 329 | }; |
| 330 | |
| 331 | // Base class for sample counter key with context |
| 332 | struct ContextKey { |
| 333 | uint64_t HashCode = 0; |
| 334 | virtual ~ContextKey() = default; |
| 335 | uint64_t getHashCode() { |
| 336 | if (HashCode == 0) |
| 337 | genHashCode(); |
| 338 | return HashCode; |
| 339 | } |
| 340 | virtual void genHashCode() = 0; |
| 341 | virtual bool isEqual(const ContextKey *K) const { |
| 342 | return HashCode == K->HashCode; |
| 343 | }; |
| 344 | |
| 345 | // Utilities for LLVM-style RTTI |
| 346 | enum ContextKind { CK_StringBased, CK_AddrBased }; |
| 347 | const ContextKind Kind; |
| 348 | ContextKind getKind() const { return Kind; } |
| 349 | ContextKey(ContextKind K) : Kind(K){}; |
| 350 | }; |
| 351 | |
| 352 | // String based context id |
| 353 | struct StringBasedCtxKey : public ContextKey { |
| 354 | SampleContextFrameVector Context; |
| 355 | |
| 356 | bool WasLeafInlined; |
| 357 | StringBasedCtxKey() : ContextKey(CK_StringBased), WasLeafInlined(false){}; |
| 358 | static bool classof(const ContextKey *K) { |
| 359 | return K->getKind() == CK_StringBased; |
| 360 | } |
| 361 | |
| 362 | bool isEqual(const ContextKey *K) const override { |
| 363 | const StringBasedCtxKey *Other = dyn_cast<StringBasedCtxKey>(Val: K); |
| 364 | return Context == Other->Context; |
| 365 | } |
| 366 | |
| 367 | void genHashCode() override { |
| 368 | HashCode = hash_value(S: SampleContextFrames(Context)); |
| 369 | } |
| 370 | }; |
| 371 | |
| 372 | // Address-based context id |
| 373 | struct AddrBasedCtxKey : public ContextKey { |
| 374 | SmallVector<uint64_t, 16> Context; |
| 375 | |
| 376 | bool WasLeafInlined; |
| 377 | AddrBasedCtxKey() : ContextKey(CK_AddrBased), WasLeafInlined(false){}; |
| 378 | static bool classof(const ContextKey *K) { |
| 379 | return K->getKind() == CK_AddrBased; |
| 380 | } |
| 381 | |
| 382 | bool isEqual(const ContextKey *K) const override { |
| 383 | const AddrBasedCtxKey *Other = dyn_cast<AddrBasedCtxKey>(Val: K); |
| 384 | return Context == Other->Context; |
| 385 | } |
| 386 | |
| 387 | void genHashCode() override { HashCode = hash_combine_range(R&: Context); } |
| 388 | }; |
| 389 | |
| 390 | // The counter of branch samples for one function indexed by the branch, |
| 391 | // which is represented as the source and target offset pair. |
| 392 | using BranchSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
| 393 | // The counter of range samples for one function indexed by the range, |
| 394 | // which is represented as the start and end offset pair. |
| 395 | using RangeSample = std::map<std::pair<uint64_t, uint64_t>, uint64_t>; |
| 396 | // <<inst-addr, vtable-data-symbol>, count> map for data access samples. |
| 397 | // The instruction address is the virtual address in the binary. |
| 398 | using DataAccessSample = std::map<std::pair<uint64_t, StringRef>, uint64_t>; |
| 399 | // Wrapper for sample counters including range counter and branch counter |
| 400 | struct SampleCounter { |
| 401 | RangeSample RangeCounter; |
| 402 | BranchSample BranchCounter; |
| 403 | DataAccessSample DataAccessCounter; |
| 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 | void recordDataAccessCount(uint64_t InstAddr, StringRef DataSymbol, |
| 413 | uint64_t Repeat) { |
| 414 | DataAccessCounter[{InstAddr, DataSymbol}] += Repeat; |
| 415 | } |
| 416 | }; |
| 417 | |
| 418 | // Sample counter with context to support context-sensitive profile |
| 419 | using ContextSampleCounterMap = |
| 420 | std::unordered_map<Hashable<ContextKey>, SampleCounter, |
| 421 | Hashable<ContextKey>::Hash, Hashable<ContextKey>::Equal>; |
| 422 | |
| 423 | struct FrameStack { |
| 424 | SmallVector<uint64_t, 16> Stack; |
| 425 | ProfiledBinary *Binary; |
| 426 | FrameStack(ProfiledBinary *B) : Binary(B) {} |
| 427 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
| 428 | assert(!Cur->isExternalFrame() && |
| 429 | "External frame's not expected for context stack." ); |
| 430 | Stack.push_back(Elt: Cur->Address); |
| 431 | return true; |
| 432 | } |
| 433 | |
| 434 | void popFrame() { |
| 435 | if (!Stack.empty()) |
| 436 | Stack.pop_back(); |
| 437 | } |
| 438 | std::shared_ptr<StringBasedCtxKey> getContextKey(); |
| 439 | }; |
| 440 | |
| 441 | struct AddressStack { |
| 442 | SmallVector<uint64_t, 16> Stack; |
| 443 | ProfiledBinary *Binary; |
| 444 | AddressStack(ProfiledBinary *B) : Binary(B) {} |
| 445 | bool pushFrame(UnwindState::ProfiledFrame *Cur) { |
| 446 | assert(!Cur->isExternalFrame() && |
| 447 | "External frame's not expected for context stack." ); |
| 448 | Stack.push_back(Elt: Cur->Address); |
| 449 | return true; |
| 450 | } |
| 451 | |
| 452 | void popFrame() { |
| 453 | if (!Stack.empty()) |
| 454 | Stack.pop_back(); |
| 455 | } |
| 456 | std::shared_ptr<AddrBasedCtxKey> getContextKey(); |
| 457 | }; |
| 458 | |
| 459 | /* |
| 460 | As in hybrid sample we have a group of LBRs and the most recent sampling call |
| 461 | stack, we can walk through those LBRs to infer more call stacks which would be |
| 462 | used as context for profile. VirtualUnwinder is the class to do the call stack |
| 463 | unwinding based on LBR state. Two types of unwinding are processd here: |
| 464 | 1) LBR unwinding and 2) linear range unwinding. |
| 465 | Specifically, for each LBR entry(can be classified into call, return, regular |
| 466 | branch), LBR unwinding will replay the operation by pushing, popping or |
| 467 | switching leaf frame towards the call stack and since the initial call stack |
| 468 | is most recently sampled, the replay should be in anti-execution order, i.e. for |
| 469 | the regular case, pop the call stack when LBR is call, push frame on call stack |
| 470 | when LBR is return. After each LBR processed, it also needs to align with the |
| 471 | next LBR by going through instructions from previous LBR's target to current |
| 472 | LBR's source, which is the linear unwinding. As instruction from linear range |
| 473 | can come from different function by inlining, linear unwinding will do the range |
| 474 | splitting and record counters by the range with same inline context. Over those |
| 475 | unwinding process we will record each call stack as context id and LBR/linear |
| 476 | range as sample counter for further CS profile generation. |
| 477 | */ |
| 478 | class VirtualUnwinder { |
| 479 | public: |
| 480 | VirtualUnwinder(ContextSampleCounterMap *Counter, ProfiledBinary *B) |
| 481 | : CtxCounterMap(Counter), Binary(B) {} |
| 482 | bool unwind(const PerfSample *Sample, uint64_t Repeat); |
| 483 | std::set<uint64_t> &getUntrackedCallsites() { return UntrackedCallsites; } |
| 484 | |
| 485 | uint64_t NumTotalBranches = 0; |
| 486 | uint64_t NumExtCallBranch = 0; |
| 487 | uint64_t NumMissingExternalFrame = 0; |
| 488 | uint64_t NumMismatchedProEpiBranch = 0; |
| 489 | uint64_t NumMismatchedExtCallBranch = 0; |
| 490 | uint64_t NumUnpairedExtAddr = 0; |
| 491 | uint64_t NumPairedExtAddr = 0; |
| 492 | |
| 493 | private: |
| 494 | bool isSourceExternal(UnwindState &State) const { |
| 495 | return State.getCurrentLBRSource() == ExternalAddr; |
| 496 | } |
| 497 | |
| 498 | bool isTargetExternal(UnwindState &State) const { |
| 499 | return State.getCurrentLBRTarget() == ExternalAddr; |
| 500 | } |
| 501 | |
| 502 | // Determine whether the return source is from external code by checking if |
| 503 | // the target's the next inst is a call inst. |
| 504 | bool isReturnFromExternal(UnwindState &State) const { |
| 505 | return isSourceExternal(State) && |
| 506 | (Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) != 0); |
| 507 | } |
| 508 | |
| 509 | // If the source is external address but it's not the `return` case, treat it |
| 510 | // as a call from external. |
| 511 | bool isCallFromExternal(UnwindState &State) const { |
| 512 | return isSourceExternal(State) && |
| 513 | Binary->getCallAddrFromFrameAddr(FrameAddr: State.getCurrentLBRTarget()) == 0; |
| 514 | } |
| 515 | |
| 516 | bool isCallState(UnwindState &State) const { |
| 517 | // The tail call frame is always missing here in stack sample, we will |
| 518 | // use a specific tail call tracker to infer it. |
| 519 | if (!isValidState(State)) |
| 520 | return false; |
| 521 | |
| 522 | if (Binary->addressIsCall(Address: State.getCurrentLBRSource())) |
| 523 | return true; |
| 524 | |
| 525 | return isCallFromExternal(State); |
| 526 | } |
| 527 | |
| 528 | bool isReturnState(UnwindState &State) const { |
| 529 | if (!isValidState(State)) |
| 530 | return false; |
| 531 | |
| 532 | // Simply check addressIsReturn, as ret is always reliable, both for |
| 533 | // regular call and tail call. |
| 534 | if (Binary->addressIsReturn(Address: State.getCurrentLBRSource())) |
| 535 | return true; |
| 536 | |
| 537 | return isReturnFromExternal(State); |
| 538 | } |
| 539 | |
| 540 | bool isValidState(UnwindState &State) const { return !State.Invalid; } |
| 541 | |
| 542 | void unwindCall(UnwindState &State); |
| 543 | void unwindLinear(UnwindState &State, uint64_t Repeat); |
| 544 | void unwindReturn(UnwindState &State); |
| 545 | void unwindBranch(UnwindState &State); |
| 546 | |
| 547 | template <typename T> |
| 548 | void collectSamplesFromFrame(UnwindState::ProfiledFrame *Cur, T &Stack); |
| 549 | // Collect each samples on trie node by DFS traversal |
| 550 | template <typename T> |
| 551 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur, T &Stack); |
| 552 | void collectSamplesFromFrameTrie(UnwindState::ProfiledFrame *Cur); |
| 553 | |
| 554 | void recordRangeCount(uint64_t Start, uint64_t End, UnwindState &State, |
| 555 | uint64_t Repeat); |
| 556 | void recordBranchCount(const LBREntry &Branch, UnwindState &State, |
| 557 | uint64_t Repeat); |
| 558 | |
| 559 | ContextSampleCounterMap *CtxCounterMap; |
| 560 | // Profiled binary that current frame address belongs to |
| 561 | ProfiledBinary *Binary; |
| 562 | // Keep track of all untracked callsites |
| 563 | std::set<uint64_t> UntrackedCallsites; |
| 564 | }; |
| 565 | |
| 566 | // Read perf trace to parse the events and samples. |
| 567 | class PerfReaderBase { |
| 568 | public: |
| 569 | PerfReaderBase(ProfiledBinary *B, StringRef PerfTrace) |
| 570 | : Binary(B), PerfTraceFile(PerfTrace) { |
| 571 | // Initialize the base address to preferred address. |
| 572 | Binary->setBaseAddress(Binary->getPreferredBaseAddress()); |
| 573 | }; |
| 574 | virtual ~PerfReaderBase() = default; |
| 575 | static std::unique_ptr<PerfReaderBase> |
| 576 | create(ProfiledBinary *Binary, PerfInputFile &PerfInput, |
| 577 | std::optional<int32_t> PIDFilter); |
| 578 | |
| 579 | // Entry of the reader to parse multiple perf traces |
| 580 | virtual void parsePerfTraces() = 0; |
| 581 | |
| 582 | // Parse the <ip, vtable-data-symbol> from the data access perf trace file, |
| 583 | // and accumulate the data access count for each <ip, data-symbol> pair. |
| 584 | Error |
| 585 | parseDataAccessPerfTraces(StringRef DataAccessPerfFile, |
| 586 | std::optional<int32_t> PIDFilter = std::nullopt); |
| 587 | |
| 588 | const ContextSampleCounterMap &getSampleCounters() const { |
| 589 | return SampleCounters; |
| 590 | } |
| 591 | bool profileIsCS() { return ProfileIsCS; } |
| 592 | |
| 593 | protected: |
| 594 | ProfiledBinary *Binary = nullptr; |
| 595 | StringRef PerfTraceFile; |
| 596 | |
| 597 | ContextSampleCounterMap SampleCounters; |
| 598 | bool ProfileIsCS = false; |
| 599 | |
| 600 | uint64_t NumTotalSample = 0; |
| 601 | uint64_t NumLeafExternalFrame = 0; |
| 602 | uint64_t NumLeadingOutgoingLBR = 0; |
| 603 | }; |
| 604 | |
| 605 | // Read perf script to parse the events and samples. |
| 606 | class PerfScriptReader : public PerfReaderBase { |
| 607 | public: |
| 608 | PerfScriptReader(ProfiledBinary *B, StringRef PerfTrace, |
| 609 | std::optional<int32_t> PID) |
| 610 | : PerfReaderBase(B, PerfTrace), PIDFilter(PID) {}; |
| 611 | |
| 612 | // Entry of the reader to parse multiple perf traces |
| 613 | void parsePerfTraces() override; |
| 614 | |
| 615 | // Parse a single line of a PERF_RECORD_MMAP event looking for a |
| 616 | // mapping between the binary name and its memory layout. |
| 617 | // TODO: Move this static method from PerScriptReader (subclass) to |
| 618 | // PerfReaderBase (superclass). |
| 619 | static bool (ProfiledBinary *Binary, StringRef Line, |
| 620 | MMapEvent &MMap); |
| 621 | |
| 622 | // Generate perf script from perf data |
| 623 | static PerfInputFile convertPerfDataToTrace(ProfiledBinary *Binary, |
| 624 | bool SkipPID, PerfInputFile &File, |
| 625 | std::optional<int32_t> PIDFilter); |
| 626 | // Extract perf script type by peaking at the input |
| 627 | static PerfContent checkPerfScriptType(StringRef FileName); |
| 628 | |
| 629 | // Cleanup installers for temporary files created by perf script command. |
| 630 | // Those files will be automatically removed when running destructor or |
| 631 | // receiving signals. |
| 632 | static SmallVector<CleanupInstaller, 2> TempFileCleanups; |
| 633 | |
| 634 | protected: |
| 635 | // Check whether a given line is LBR sample |
| 636 | static bool isLBRSample(StringRef Line); |
| 637 | // Check whether a given line is MMAP event |
| 638 | static bool isMMapEvent(StringRef Line); |
| 639 | // Update base address based on mmap events |
| 640 | void updateBinaryAddress(const MMapEvent &Event); |
| 641 | // Parse mmap event and update binary address |
| 642 | void parseMMapEvent(TraceStream &TraceIt); |
| 643 | // Parse perf events/samples and do aggregation |
| 644 | void parseAndAggregateTrace(); |
| 645 | // Parse either an MMAP event or a perf sample |
| 646 | void parseEventOrSample(TraceStream &TraceIt); |
| 647 | // Warn if the relevant mmap event is missing. |
| 648 | void warnIfMissingMMap(); |
| 649 | // Emit accumulate warnings. |
| 650 | void warnTruncatedStack(); |
| 651 | // Warn if range is invalid. |
| 652 | void warnInvalidRange(); |
| 653 | // Extract call stack from the perf trace lines |
| 654 | bool (TraceStream &TraceIt, |
| 655 | SmallVectorImpl<uint64_t> &CallStack); |
| 656 | // Extract LBR stack from one perf trace line |
| 657 | bool (TraceStream &TraceIt, |
| 658 | SmallVectorImpl<LBREntry> &LBRStack); |
| 659 | uint64_t parseAggregatedCount(TraceStream &TraceIt); |
| 660 | // Parse one sample from multiple perf lines, override this for different |
| 661 | // sample type |
| 662 | void parseSample(TraceStream &TraceIt); |
| 663 | // An aggregated count is given to indicate how many times the sample is |
| 664 | // repeated. |
| 665 | virtual void parseSample(TraceStream &TraceIt, uint64_t Count){}; |
| 666 | void computeCounterFromLBR(const PerfSample *Sample, uint64_t Repeat); |
| 667 | // Post process the profile after trace aggregation, we will do simple range |
| 668 | // overlap computation for AutoFDO, or unwind for CSSPGO(hybrid sample). |
| 669 | virtual void generateUnsymbolizedProfile(); |
| 670 | void writeUnsymbolizedProfile(StringRef Filename); |
| 671 | void writeUnsymbolizedProfile(raw_fd_ostream &OS); |
| 672 | |
| 673 | // Samples with the repeating time generated by the perf reader |
| 674 | AggregatedCounter AggregatedSamples; |
| 675 | // Keep track of all invalid return addresses |
| 676 | std::set<uint64_t> InvalidReturnAddresses; |
| 677 | // PID for the process of interest |
| 678 | std::optional<int32_t> PIDFilter; |
| 679 | }; |
| 680 | |
| 681 | /* |
| 682 | The reader of LBR only perf script. |
| 683 | A typical LBR sample is like: |
| 684 | 40062f 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
| 685 | ... 0x4005c8/0x4005dc/P/-/-/0 |
| 686 | */ |
| 687 | class LBRPerfReader : public PerfScriptReader { |
| 688 | public: |
| 689 | LBRPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
| 690 | std::optional<int32_t> PID) |
| 691 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
| 692 | // Parse the LBR only sample. |
| 693 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
| 694 | }; |
| 695 | |
| 696 | /* |
| 697 | Hybrid perf script includes a group of hybrid samples(LBRs + call stack), |
| 698 | which is used to generate CS profile. An example of hybrid sample: |
| 699 | 4005dc # call stack leaf |
| 700 | 400634 |
| 701 | 400684 # call stack root |
| 702 | 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... |
| 703 | ... 0x4005c8/0x4005dc/P/-/-/0 # LBR Entries |
| 704 | */ |
| 705 | class HybridPerfReader : public PerfScriptReader { |
| 706 | public: |
| 707 | HybridPerfReader(ProfiledBinary *Binary, StringRef PerfTrace, |
| 708 | std::optional<int32_t> PID) |
| 709 | : PerfScriptReader(Binary, PerfTrace, PID) {}; |
| 710 | // Parse the hybrid sample including the call and LBR line |
| 711 | void parseSample(TraceStream &TraceIt, uint64_t Count) override; |
| 712 | void generateUnsymbolizedProfile() override; |
| 713 | |
| 714 | private: |
| 715 | // Unwind the hybrid samples after aggregration |
| 716 | void unwindSamples(); |
| 717 | }; |
| 718 | |
| 719 | /* |
| 720 | Format of unsymbolized profile: |
| 721 | |
| 722 | [frame1 @ frame2 @ ...] # If it's a CS profile |
| 723 | number of entries in RangeCounter |
| 724 | from_1-to_1:count_1 |
| 725 | from_2-to_2:count_2 |
| 726 | ...... |
| 727 | from_n-to_n:count_n |
| 728 | number of entries in BranchCounter |
| 729 | src_1->dst_1:count_1 |
| 730 | src_2->dst_2:count_2 |
| 731 | ...... |
| 732 | src_n->dst_n:count_n |
| 733 | [frame1 @ frame2 @ ...] # Next context |
| 734 | ...... |
| 735 | |
| 736 | Note that non-CS profile doesn't have the empty `[]` context. |
| 737 | */ |
| 738 | class UnsymbolizedProfileReader : public PerfReaderBase { |
| 739 | public: |
| 740 | UnsymbolizedProfileReader(ProfiledBinary *Binary, StringRef PerfTrace) |
| 741 | : PerfReaderBase(Binary, PerfTrace){}; |
| 742 | void parsePerfTraces() override; |
| 743 | |
| 744 | private: |
| 745 | void readSampleCounters(TraceStream &TraceIt, SampleCounter &SCounters); |
| 746 | void readUnsymbolizedProfile(StringRef Filename); |
| 747 | |
| 748 | std::unordered_set<std::string> ContextStrSet; |
| 749 | }; |
| 750 | |
| 751 | } // end namespace sampleprof |
| 752 | } // end namespace llvm |
| 753 | |
| 754 | #endif |
| 755 | |