1 //===-- PerfReader.cpp - 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 #include "PerfReader.h" 9 #include "ProfileGenerator.h" 10 #include "llvm/DebugInfo/Symbolize/SymbolizableModule.h" 11 #include "llvm/Support/FileSystem.h" 12 #include "llvm/Support/Process.h" 13 14 #define DEBUG_TYPE "perf-reader" 15 16 cl::opt<bool> SkipSymbolization("skip-symbolization", cl::init(false), 17 cl::ZeroOrMore, 18 cl::desc("Dump the unsymbolized profile to the " 19 "output file. It will show unwinder " 20 "output for CS profile generation.")); 21 22 static cl::opt<bool> ShowMmapEvents("show-mmap-events", cl::init(false), 23 cl::ZeroOrMore, 24 cl::desc("Print binary load events.")); 25 26 static cl::opt<bool> 27 UseOffset("use-offset", cl::init(true), cl::ZeroOrMore, 28 cl::desc("Work with `--skip-symbolization` or " 29 "`--unsymbolized-profile` to write/read the " 30 "offset instead of virtual address.")); 31 32 static cl::opt<bool> UseLoadableSegmentAsBase( 33 "use-first-loadable-segment-as-base", cl::init(false), cl::ZeroOrMore, 34 cl::desc("Use first loadable segment address as base address " 35 "for offsets in unsymbolized profile. By default " 36 "first executable segment address is used")); 37 38 static cl::opt<bool> 39 IgnoreStackSamples("ignore-stack-samples", cl::init(false), cl::ZeroOrMore, 40 cl::desc("Ignore call stack samples for hybrid samples " 41 "and produce context-insensitive profile.")); 42 cl::opt<bool> ShowDetailedWarning("show-detailed-warning", cl::init(false), 43 cl::ZeroOrMore, 44 cl::desc("Show detailed warning message.")); 45 46 extern cl::opt<std::string> PerfTraceFilename; 47 extern cl::opt<bool> ShowDisassemblyOnly; 48 extern cl::opt<bool> ShowSourceLocations; 49 extern cl::opt<std::string> OutputFilename; 50 51 namespace llvm { 52 namespace sampleprof { 53 54 void VirtualUnwinder::unwindCall(UnwindState &State) { 55 uint64_t Source = State.getCurrentLBRSource(); 56 // An artificial return should push an external frame and an artificial call 57 // will match it and pop the external frame so that the context before and 58 // after the external call will be the same. 59 if (State.getCurrentLBR().IsArtificial) { 60 NumExtCallBranch++; 61 // A return is matched and pop the external frame. 62 if (State.getParentFrame()->isExternalFrame()) { 63 State.popFrame(); 64 } else { 65 // An artificial return is missing, it happens that the sample is just hit 66 // in the middle of the external code. In this case, the leading branch is 67 // a call to external, we just keep unwinding use a context-less stack. 68 if (State.getParentFrame() != State.getDummyRootPtr()) 69 NumMissingExternalFrame++; 70 State.clearCallStack(); 71 State.pushFrame(Source); 72 State.InstPtr.update(Source); 73 return; 74 } 75 } 76 77 auto *ParentFrame = State.getParentFrame(); 78 // The 2nd frame after leaf could be missing if stack sample is 79 // taken when IP is within prolog/epilog, as frame chain isn't 80 // setup yet. Fill in the missing frame in that case. 81 // TODO: Currently we just assume all the addr that can't match the 82 // 2nd frame is in prolog/epilog. In the future, we will switch to 83 // pro/epi tracker(Dwarf CFI) for the precise check. 84 if (ParentFrame == State.getDummyRootPtr() || 85 ParentFrame->Address != Source) { 86 State.switchToFrame(Source); 87 if (ParentFrame != State.getDummyRootPtr()) { 88 if (State.getCurrentLBR().IsArtificial) 89 NumMismatchedExtCallBranch++; 90 else 91 NumMismatchedProEpiBranch++; 92 } 93 } else { 94 State.popFrame(); 95 } 96 State.InstPtr.update(Source); 97 } 98 99 void VirtualUnwinder::unwindLinear(UnwindState &State, uint64_t Repeat) { 100 InstructionPointer &IP = State.InstPtr; 101 uint64_t Target = State.getCurrentLBRTarget(); 102 uint64_t End = IP.Address; 103 if (Target > End) { 104 // Skip unwinding the rest of LBR trace when a bogus range is seen. 105 State.setInvalid(); 106 return; 107 } 108 if (Binary->usePseudoProbes()) { 109 // We don't need to top frame probe since it should be extracted 110 // from the range. 111 // The outcome of the virtual unwinding with pseudo probes is a 112 // map from a context key to the address range being unwound. 113 // This means basically linear unwinding is not needed for pseudo 114 // probes. The range will be simply recorded here and will be 115 // converted to a list of pseudo probes to report in ProfileGenerator. 116 State.getParentFrame()->recordRangeCount(Target, End, Repeat); 117 } else { 118 // Unwind linear execution part. 119 // Split and record the range by different inline context. For example: 120 // [0x01] ... main:1 # Target 121 // [0x02] ... main:2 122 // [0x03] ... main:3 @ foo:1 123 // [0x04] ... main:3 @ foo:2 124 // [0x05] ... main:3 @ foo:3 125 // [0x06] ... main:4 126 // [0x07] ... main:5 # End 127 // It will be recorded: 128 // [main:*] : [0x06, 0x07], [0x01, 0x02] 129 // [main:3 @ foo:*] : [0x03, 0x05] 130 while (IP.Address > Target) { 131 uint64_t PrevIP = IP.Address; 132 IP.backward(); 133 // Break into segments for implicit call/return due to inlining 134 bool SameInlinee = Binary->inlineContextEqual(PrevIP, IP.Address); 135 if (!SameInlinee) { 136 State.switchToFrame(PrevIP); 137 State.CurrentLeafFrame->recordRangeCount(PrevIP, End, Repeat); 138 End = IP.Address; 139 } 140 } 141 assert(IP.Address == Target && "The last one must be the target address."); 142 // Record the remaining range, [0x01, 0x02] in the example 143 State.switchToFrame(IP.Address); 144 State.CurrentLeafFrame->recordRangeCount(IP.Address, End, Repeat); 145 } 146 } 147 148 void VirtualUnwinder::unwindReturn(UnwindState &State) { 149 // Add extra frame as we unwind through the return 150 const LBREntry &LBR = State.getCurrentLBR(); 151 uint64_t CallAddr = Binary->getCallAddrFromFrameAddr(LBR.Target); 152 State.switchToFrame(CallAddr); 153 // Push an external frame for the case of returning to external 154 // address(callback), later if an aitificial call is matched and it will be 155 // popped up. This is to 1)avoid context being interrupted by callback, 156 // context before or after the callback should be the same. 2) the call stack 157 // of function called by callback should be truncated which is done during 158 // recording the context on trie. For example: 159 // main (call)--> foo (call)--> callback (call)--> bar (return)--> callback 160 // (return)--> foo (return)--> main 161 // Context for bar should not include main and foo. 162 // For the code of foo, the context of before and after callback should both 163 // be [foo, main]. 164 if (LBR.IsArtificial) 165 State.pushFrame(ExternalAddr); 166 State.pushFrame(LBR.Source); 167 State.InstPtr.update(LBR.Source); 168 } 169 170 void VirtualUnwinder::unwindBranch(UnwindState &State) { 171 // TODO: Tolerate tail call for now, as we may see tail call from libraries. 172 // This is only for intra function branches, excluding tail calls. 173 uint64_t Source = State.getCurrentLBRSource(); 174 State.switchToFrame(Source); 175 State.InstPtr.update(Source); 176 } 177 178 std::shared_ptr<StringBasedCtxKey> FrameStack::getContextKey() { 179 std::shared_ptr<StringBasedCtxKey> KeyStr = 180 std::make_shared<StringBasedCtxKey>(); 181 KeyStr->Context = Binary->getExpandedContext(Stack, KeyStr->WasLeafInlined); 182 if (KeyStr->Context.empty()) 183 return nullptr; 184 return KeyStr; 185 } 186 187 std::shared_ptr<AddrBasedCtxKey> AddressStack::getContextKey() { 188 std::shared_ptr<AddrBasedCtxKey> KeyStr = std::make_shared<AddrBasedCtxKey>(); 189 KeyStr->Context = Stack; 190 CSProfileGenerator::compressRecursionContext<uint64_t>(KeyStr->Context); 191 CSProfileGenerator::trimContext<uint64_t>(KeyStr->Context); 192 return KeyStr; 193 } 194 195 template <typename T> 196 void VirtualUnwinder::collectSamplesFromFrame(UnwindState::ProfiledFrame *Cur, 197 T &Stack) { 198 if (Cur->RangeSamples.empty() && Cur->BranchSamples.empty()) 199 return; 200 201 std::shared_ptr<ContextKey> Key = Stack.getContextKey(); 202 if (Key == nullptr) 203 return; 204 auto Ret = CtxCounterMap->emplace(Hashable<ContextKey>(Key), SampleCounter()); 205 SampleCounter &SCounter = Ret.first->second; 206 for (auto &Item : Cur->RangeSamples) { 207 uint64_t StartOffset = Binary->virtualAddrToOffset(std::get<0>(Item)); 208 uint64_t EndOffset = Binary->virtualAddrToOffset(std::get<1>(Item)); 209 SCounter.recordRangeCount(StartOffset, EndOffset, std::get<2>(Item)); 210 } 211 212 for (auto &Item : Cur->BranchSamples) { 213 uint64_t SourceOffset = Binary->virtualAddrToOffset(std::get<0>(Item)); 214 uint64_t TargetOffset = Binary->virtualAddrToOffset(std::get<1>(Item)); 215 SCounter.recordBranchCount(SourceOffset, TargetOffset, std::get<2>(Item)); 216 } 217 } 218 219 template <typename T> 220 void VirtualUnwinder::collectSamplesFromFrameTrie( 221 UnwindState::ProfiledFrame *Cur, T &Stack) { 222 if (!Cur->isDummyRoot()) { 223 // Truncate the context for external frame since this isn't a real call 224 // context the compiler will see. 225 if (Cur->isExternalFrame() || !Stack.pushFrame(Cur)) { 226 // Process truncated context 227 // Start a new traversal ignoring its bottom context 228 T EmptyStack(Binary); 229 collectSamplesFromFrame(Cur, EmptyStack); 230 for (const auto &Item : Cur->Children) { 231 collectSamplesFromFrameTrie(Item.second.get(), EmptyStack); 232 } 233 234 // Keep note of untracked call site and deduplicate them 235 // for warning later. 236 if (!Cur->isLeafFrame()) 237 UntrackedCallsites.insert(Cur->Address); 238 239 return; 240 } 241 } 242 243 collectSamplesFromFrame(Cur, Stack); 244 // Process children frame 245 for (const auto &Item : Cur->Children) { 246 collectSamplesFromFrameTrie(Item.second.get(), Stack); 247 } 248 // Recover the call stack 249 Stack.popFrame(); 250 } 251 252 void VirtualUnwinder::collectSamplesFromFrameTrie( 253 UnwindState::ProfiledFrame *Cur) { 254 if (Binary->usePseudoProbes()) { 255 AddressStack Stack(Binary); 256 collectSamplesFromFrameTrie<AddressStack>(Cur, Stack); 257 } else { 258 FrameStack Stack(Binary); 259 collectSamplesFromFrameTrie<FrameStack>(Cur, Stack); 260 } 261 } 262 263 void VirtualUnwinder::recordBranchCount(const LBREntry &Branch, 264 UnwindState &State, uint64_t Repeat) { 265 if (Branch.IsArtificial || Branch.Target == ExternalAddr) 266 return; 267 268 if (Binary->usePseudoProbes()) { 269 // Same as recordRangeCount, We don't need to top frame probe since we will 270 // extract it from branch's source address 271 State.getParentFrame()->recordBranchCount(Branch.Source, Branch.Target, 272 Repeat); 273 } else { 274 State.CurrentLeafFrame->recordBranchCount(Branch.Source, Branch.Target, 275 Repeat); 276 } 277 } 278 279 bool VirtualUnwinder::unwind(const PerfSample *Sample, uint64_t Repeat) { 280 // Capture initial state as starting point for unwinding. 281 UnwindState State(Sample, Binary); 282 283 // Sanity check - making sure leaf of LBR aligns with leaf of stack sample 284 // Stack sample sometimes can be unreliable, so filter out bogus ones. 285 if (!State.validateInitialState()) 286 return false; 287 288 // Now process the LBR samples in parrallel with stack sample 289 // Note that we do not reverse the LBR entry order so we can 290 // unwind the sample stack as we walk through LBR entries. 291 while (State.hasNextLBR()) { 292 State.checkStateConsistency(); 293 294 // Do not attempt linear unwind for the leaf range as it's incomplete. 295 if (!State.IsLastLBR()) { 296 // Unwind implicit calls/returns from inlining, along the linear path, 297 // break into smaller sub section each with its own calling context. 298 unwindLinear(State, Repeat); 299 } 300 301 // Save the LBR branch before it gets unwound. 302 const LBREntry &Branch = State.getCurrentLBR(); 303 304 if (isCallState(State)) { 305 // Unwind calls - we know we encountered call if LBR overlaps with 306 // transition between leaf the 2nd frame. Note that for calls that 307 // were not in the original stack sample, we should have added the 308 // extra frame when processing the return paired with this call. 309 unwindCall(State); 310 } else if (isReturnState(State)) { 311 // Unwind returns - check whether the IP is indeed at a return 312 // instruction 313 unwindReturn(State); 314 } else if (isValidState(State)) { 315 // Unwind branches 316 // For regular intra function branches, we only need to record branch 317 // with context. For an artificial branch cross function boundaries, we 318 // got an issue with returning to external code. Take the two LBR enties 319 // for example: [foo:8(RETURN), ext:1] [ext:3(CALL), bar:1] After perf 320 // reader, we only get[foo:8(RETURN), bar:1], unwinder will be confused 321 // like foo return to bar. Here we detect and treat this case as BRANCH 322 // instead of RETURN which only update the source address. 323 unwindBranch(State); 324 } else { 325 // Skip unwinding the rest of LBR trace. Reset the stack and update the 326 // state so that the rest of the trace can still be processed as if they 327 // do not have stack samples. 328 State.clearCallStack(); 329 State.InstPtr.update(State.getCurrentLBRSource()); 330 State.pushFrame(State.InstPtr.Address); 331 } 332 333 State.advanceLBR(); 334 // Record `branch` with calling context after unwinding. 335 recordBranchCount(Branch, State, Repeat); 336 } 337 // As samples are aggregated on trie, record them into counter map 338 collectSamplesFromFrameTrie(State.getDummyRootPtr()); 339 340 return true; 341 } 342 343 std::unique_ptr<PerfReaderBase> 344 PerfReaderBase::create(ProfiledBinary *Binary, PerfInputFile &PerfInput, 345 Optional<uint32_t> PIDFilter) { 346 std::unique_ptr<PerfReaderBase> PerfReader; 347 348 if (PerfInput.Format == PerfFormat::UnsymbolizedProfile) { 349 PerfReader.reset( 350 new UnsymbolizedProfileReader(Binary, PerfInput.InputFile)); 351 return PerfReader; 352 } 353 354 // For perf data input, we need to convert them into perf script first. 355 if (PerfInput.Format == PerfFormat::PerfData) 356 PerfInput = 357 PerfScriptReader::convertPerfDataToTrace(Binary, PerfInput, PIDFilter); 358 359 assert((PerfInput.Format == PerfFormat::PerfScript) && 360 "Should be a perfscript!"); 361 362 PerfInput.Content = 363 PerfScriptReader::checkPerfScriptType(PerfInput.InputFile); 364 if (PerfInput.Content == PerfContent::LBRStack) { 365 PerfReader.reset( 366 new HybridPerfReader(Binary, PerfInput.InputFile, PIDFilter)); 367 } else if (PerfInput.Content == PerfContent::LBR) { 368 PerfReader.reset(new LBRPerfReader(Binary, PerfInput.InputFile, PIDFilter)); 369 } else { 370 exitWithError("Unsupported perfscript!"); 371 } 372 373 return PerfReader; 374 } 375 376 PerfInputFile PerfScriptReader::convertPerfDataToTrace( 377 ProfiledBinary *Binary, PerfInputFile &File, Optional<uint32_t> PIDFilter) { 378 StringRef PerfData = File.InputFile; 379 // Run perf script to retrieve PIDs matching binary we're interested in. 380 auto PerfExecutable = sys::Process::FindInEnvPath("PATH", "perf"); 381 if (!PerfExecutable) { 382 exitWithError("Perf not found."); 383 } 384 std::string PerfPath = *PerfExecutable; 385 std::string PerfTraceFile = PerfData.str() + ".script.tmp"; 386 StringRef ScriptMMapArgs[] = {PerfPath, "script", "--show-mmap-events", 387 "-F", "comm,pid", "-i", 388 PerfData}; 389 Optional<StringRef> Redirects[] = {llvm::None, // Stdin 390 StringRef(PerfTraceFile), // Stdout 391 StringRef(PerfTraceFile)}; // Stderr 392 sys::ExecuteAndWait(PerfPath, ScriptMMapArgs, llvm::None, Redirects); 393 394 // Collect the PIDs 395 TraceStream TraceIt(PerfTraceFile); 396 std::string PIDs; 397 std::unordered_set<uint32_t> PIDSet; 398 while (!TraceIt.isAtEoF()) { 399 MMapEvent MMap; 400 if (isMMap2Event(TraceIt.getCurrentLine()) && 401 extractMMap2EventForBinary(Binary, TraceIt.getCurrentLine(), MMap)) { 402 auto It = PIDSet.emplace(MMap.PID); 403 if (It.second && (!PIDFilter || MMap.PID == *PIDFilter)) { 404 if (!PIDs.empty()) { 405 PIDs.append(","); 406 } 407 PIDs.append(utostr(MMap.PID)); 408 } 409 } 410 TraceIt.advance(); 411 } 412 413 if (PIDs.empty()) { 414 exitWithError("No relevant mmap event is found in perf data."); 415 } 416 417 // Run perf script again to retrieve events for PIDs collected above 418 StringRef ScriptSampleArgs[] = {PerfPath, "script", "--show-mmap-events", 419 "-F", "ip,brstack", "--pid", 420 PIDs, "-i", PerfData}; 421 sys::ExecuteAndWait(PerfPath, ScriptSampleArgs, llvm::None, Redirects); 422 423 return {PerfTraceFile, PerfFormat::PerfScript, PerfContent::UnknownContent}; 424 } 425 426 void PerfScriptReader::updateBinaryAddress(const MMapEvent &Event) { 427 // Drop the event which doesn't belong to user-provided binary 428 StringRef BinaryName = llvm::sys::path::filename(Event.BinaryPath); 429 if (Binary->getName() != BinaryName) 430 return; 431 432 // Drop the event if process does not match pid filter 433 if (PIDFilter && Event.PID != *PIDFilter) 434 return; 435 436 // Drop the event if its image is loaded at the same address 437 if (Event.Address == Binary->getBaseAddress()) { 438 Binary->setIsLoadedByMMap(true); 439 return; 440 } 441 442 if (Event.Offset == Binary->getTextSegmentOffset()) { 443 // A binary image could be unloaded and then reloaded at different 444 // place, so update binary load address. 445 // Only update for the first executable segment and assume all other 446 // segments are loaded at consecutive memory addresses, which is the case on 447 // X64. 448 Binary->setBaseAddress(Event.Address); 449 Binary->setIsLoadedByMMap(true); 450 } else { 451 // Verify segments are loaded consecutively. 452 const auto &Offsets = Binary->getTextSegmentOffsets(); 453 auto It = std::lower_bound(Offsets.begin(), Offsets.end(), Event.Offset); 454 if (It != Offsets.end() && *It == Event.Offset) { 455 // The event is for loading a separate executable segment. 456 auto I = std::distance(Offsets.begin(), It); 457 const auto &PreferredAddrs = Binary->getPreferredTextSegmentAddresses(); 458 if (PreferredAddrs[I] - Binary->getPreferredBaseAddress() != 459 Event.Address - Binary->getBaseAddress()) 460 exitWithError("Executable segments not loaded consecutively"); 461 } else { 462 if (It == Offsets.begin()) 463 exitWithError("File offset not found"); 464 else { 465 // Find the segment the event falls in. A large segment could be loaded 466 // via multiple mmap calls with consecutive memory addresses. 467 --It; 468 assert(*It < Event.Offset); 469 if (Event.Offset - *It != Event.Address - Binary->getBaseAddress()) 470 exitWithError("Segment not loaded by consecutive mmaps"); 471 } 472 } 473 } 474 } 475 476 static std::string getContextKeyStr(ContextKey *K, 477 const ProfiledBinary *Binary) { 478 if (const auto *CtxKey = dyn_cast<StringBasedCtxKey>(K)) { 479 return SampleContext::getContextString(CtxKey->Context); 480 } else if (const auto *CtxKey = dyn_cast<AddrBasedCtxKey>(K)) { 481 std::ostringstream OContextStr; 482 for (uint32_t I = 0; I < CtxKey->Context.size(); I++) { 483 if (OContextStr.str().size()) 484 OContextStr << " @ "; 485 OContextStr << "0x" 486 << to_hexString( 487 Binary->virtualAddrToOffset(CtxKey->Context[I]), 488 false); 489 } 490 return OContextStr.str(); 491 } else { 492 llvm_unreachable("unexpected key type"); 493 } 494 } 495 496 void HybridPerfReader::unwindSamples() { 497 if (Binary->useFSDiscriminator()) 498 exitWithError("FS discriminator is not supported in CS profile."); 499 VirtualUnwinder Unwinder(&SampleCounters, Binary); 500 for (const auto &Item : AggregatedSamples) { 501 const PerfSample *Sample = Item.first.getPtr(); 502 Unwinder.unwind(Sample, Item.second); 503 } 504 505 // Warn about untracked frames due to missing probes. 506 if (ShowDetailedWarning) { 507 for (auto Address : Unwinder.getUntrackedCallsites()) 508 WithColor::warning() << "Profile context truncated due to missing probe " 509 << "for call instruction at " 510 << format("0x%" PRIx64, Address) << "\n"; 511 } 512 513 emitWarningSummary(Unwinder.getUntrackedCallsites().size(), 514 SampleCounters.size(), 515 "of profiled contexts are truncated due to missing probe " 516 "for call instruction."); 517 518 emitWarningSummary( 519 Unwinder.NumMismatchedExtCallBranch, Unwinder.NumTotalBranches, 520 "of branches'source is a call instruction but doesn't match call frame " 521 "stack, likely due to unwinding error of external frame."); 522 523 emitWarningSummary( 524 Unwinder.NumMismatchedProEpiBranch, Unwinder.NumTotalBranches, 525 "of branches'source is a call instruction but doesn't match call frame " 526 "stack, likely due to frame in prolog/epilog."); 527 528 emitWarningSummary(Unwinder.NumMissingExternalFrame, 529 Unwinder.NumExtCallBranch, 530 "of artificial call branches but doesn't have an external " 531 "frame to match."); 532 } 533 534 bool PerfScriptReader::extractLBRStack(TraceStream &TraceIt, 535 SmallVectorImpl<LBREntry> &LBRStack) { 536 // The raw format of LBR stack is like: 537 // 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... 538 // ... 0x4005c8/0x4005dc/P/-/-/0 539 // It's in FIFO order and seperated by whitespace. 540 SmallVector<StringRef, 32> Records; 541 TraceIt.getCurrentLine().split(Records, " ", -1, false); 542 auto WarnInvalidLBR = [](TraceStream &TraceIt) { 543 WithColor::warning() << "Invalid address in LBR record at line " 544 << TraceIt.getLineNumber() << ": " 545 << TraceIt.getCurrentLine() << "\n"; 546 }; 547 548 // Skip the leading instruction pointer. 549 size_t Index = 0; 550 uint64_t LeadingAddr; 551 if (!Records.empty() && !Records[0].contains('/')) { 552 if (Records[0].getAsInteger(16, LeadingAddr)) { 553 WarnInvalidLBR(TraceIt); 554 TraceIt.advance(); 555 return false; 556 } 557 Index = 1; 558 } 559 // Now extract LBR samples - note that we do not reverse the 560 // LBR entry order so we can unwind the sample stack as we walk 561 // through LBR entries. 562 uint64_t PrevTrDst = 0; 563 564 while (Index < Records.size()) { 565 auto &Token = Records[Index++]; 566 if (Token.size() == 0) 567 continue; 568 569 SmallVector<StringRef, 8> Addresses; 570 Token.split(Addresses, "/"); 571 uint64_t Src; 572 uint64_t Dst; 573 574 // Stop at broken LBR records. 575 if (Addresses.size() < 2 || Addresses[0].substr(2).getAsInteger(16, Src) || 576 Addresses[1].substr(2).getAsInteger(16, Dst)) { 577 WarnInvalidLBR(TraceIt); 578 break; 579 } 580 581 bool SrcIsInternal = Binary->addressIsCode(Src); 582 bool DstIsInternal = Binary->addressIsCode(Dst); 583 bool IsExternal = !SrcIsInternal && !DstIsInternal; 584 bool IsIncoming = !SrcIsInternal && DstIsInternal; 585 bool IsOutgoing = SrcIsInternal && !DstIsInternal; 586 bool IsArtificial = false; 587 588 // Ignore branches outside the current binary. 589 if (IsExternal) { 590 if (!PrevTrDst && !LBRStack.empty()) { 591 WithColor::warning() 592 << "Invalid transfer to external code in LBR record at line " 593 << TraceIt.getLineNumber() << ": " << TraceIt.getCurrentLine() 594 << "\n"; 595 } 596 // Do not ignore the entire samples, the remaining LBR can still be 597 // unwound using a context-less stack. 598 continue; 599 } 600 601 if (IsOutgoing) { 602 if (!PrevTrDst) { 603 // This is a leading outgoing LBR, we should keep processing the LBRs. 604 if (LBRStack.empty()) { 605 NumLeadingOutgoingLBR++; 606 // Record this LBR since current source and next LBR' target is still 607 // a valid range. 608 LBRStack.emplace_back(LBREntry(Src, ExternalAddr, false)); 609 continue; 610 } 611 // This is middle unpaired outgoing jump which is likely due to 612 // interrupt or incomplete LBR trace. Ignore current and subsequent 613 // entries since they are likely in different contexts. 614 break; 615 } 616 617 // For transition to external code, group the Source with the next 618 // availabe transition target. 619 Dst = PrevTrDst; 620 PrevTrDst = 0; 621 IsArtificial = true; 622 } else { 623 if (PrevTrDst) { 624 // If we have seen an incoming transition from external code to internal 625 // code, but not a following outgoing transition, the incoming 626 // transition is likely due to interrupt which is usually unpaired. 627 // Ignore current and subsequent entries since they are likely in 628 // different contexts. 629 break; 630 } 631 632 if (IsIncoming) { 633 // For transition from external code (such as dynamic libraries) to 634 // the current binary, keep track of the branch target which will be 635 // grouped with the Source of the last transition from the current 636 // binary. 637 PrevTrDst = Dst; 638 continue; 639 } 640 } 641 642 // TODO: filter out buggy duplicate branches on Skylake 643 644 LBRStack.emplace_back(LBREntry(Src, Dst, IsArtificial)); 645 } 646 TraceIt.advance(); 647 return !LBRStack.empty(); 648 } 649 650 bool PerfScriptReader::extractCallstack(TraceStream &TraceIt, 651 SmallVectorImpl<uint64_t> &CallStack) { 652 // The raw format of call stack is like: 653 // 4005dc # leaf frame 654 // 400634 655 // 400684 # root frame 656 // It's in bottom-up order with each frame in one line. 657 658 // Extract stack frames from sample 659 while (!TraceIt.isAtEoF() && !TraceIt.getCurrentLine().startswith(" 0x")) { 660 StringRef FrameStr = TraceIt.getCurrentLine().ltrim(); 661 uint64_t FrameAddr = 0; 662 if (FrameStr.getAsInteger(16, FrameAddr)) { 663 // We might parse a non-perf sample line like empty line and comments, 664 // skip it 665 TraceIt.advance(); 666 return false; 667 } 668 TraceIt.advance(); 669 // Currently intermixed frame from different binaries is not supported. 670 if (!Binary->addressIsCode(FrameAddr)) { 671 if (CallStack.empty()) 672 NumLeafExternalFrame++; 673 // Push a special value(ExternalAddr) for the external frames so that 674 // unwinder can still work on this with artificial Call/Return branch. 675 // After unwinding, the context will be truncated for external frame. 676 // Also deduplicate the consecutive external addresses. 677 if (CallStack.empty() || CallStack.back() != ExternalAddr) 678 CallStack.emplace_back(ExternalAddr); 679 continue; 680 } 681 682 // We need to translate return address to call address for non-leaf frames. 683 if (!CallStack.empty()) { 684 auto CallAddr = Binary->getCallAddrFromFrameAddr(FrameAddr); 685 if (!CallAddr) { 686 // Stop at an invalid return address caused by bad unwinding. This could 687 // happen to frame-pointer-based unwinding and the callee functions that 688 // do not have the frame pointer chain set up. 689 InvalidReturnAddresses.insert(FrameAddr); 690 break; 691 } 692 FrameAddr = CallAddr; 693 } 694 695 CallStack.emplace_back(FrameAddr); 696 } 697 698 // Strip out the bottom external addr. 699 if (CallStack.size() > 1 && CallStack.back() == ExternalAddr) 700 CallStack.pop_back(); 701 702 // Skip other unrelated line, find the next valid LBR line 703 // Note that even for empty call stack, we should skip the address at the 704 // bottom, otherwise the following pass may generate a truncated callstack 705 while (!TraceIt.isAtEoF() && !TraceIt.getCurrentLine().startswith(" 0x")) { 706 TraceIt.advance(); 707 } 708 // Filter out broken stack sample. We may not have complete frame info 709 // if sample end up in prolog/epilog, the result is dangling context not 710 // connected to entry point. This should be relatively rare thus not much 711 // impact on overall profile quality. However we do want to filter them 712 // out to reduce the number of different calling contexts. One instance 713 // of such case - when sample landed in prolog/epilog, somehow stack 714 // walking will be broken in an unexpected way that higher frames will be 715 // missing. 716 return !CallStack.empty() && 717 !Binary->addressInPrologEpilog(CallStack.front()); 718 } 719 720 void PerfScriptReader::warnIfMissingMMap() { 721 if (!Binary->getMissingMMapWarned() && !Binary->getIsLoadedByMMap()) { 722 WithColor::warning() << "No relevant mmap event is matched for " 723 << Binary->getName() 724 << ", will use preferred address (" 725 << format("0x%" PRIx64, 726 Binary->getPreferredBaseAddress()) 727 << ") as the base loading address!\n"; 728 // Avoid redundant warning, only warn at the first unmatched sample. 729 Binary->setMissingMMapWarned(true); 730 } 731 } 732 733 void HybridPerfReader::parseSample(TraceStream &TraceIt, uint64_t Count) { 734 // The raw hybird sample started with call stack in FILO order and followed 735 // intermediately by LBR sample 736 // e.g. 737 // 4005dc # call stack leaf 738 // 400634 739 // 400684 # call stack root 740 // 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... 741 // ... 0x4005c8/0x4005dc/P/-/-/0 # LBR Entries 742 // 743 std::shared_ptr<PerfSample> Sample = std::make_shared<PerfSample>(); 744 #ifndef NDEBUG 745 Sample->Linenum = TraceIt.getLineNumber(); 746 #endif 747 // Parsing call stack and populate into PerfSample.CallStack 748 if (!extractCallstack(TraceIt, Sample->CallStack)) { 749 // Skip the next LBR line matched current call stack 750 if (!TraceIt.isAtEoF() && TraceIt.getCurrentLine().startswith(" 0x")) 751 TraceIt.advance(); 752 return; 753 } 754 755 warnIfMissingMMap(); 756 757 if (!TraceIt.isAtEoF() && TraceIt.getCurrentLine().startswith(" 0x")) { 758 // Parsing LBR stack and populate into PerfSample.LBRStack 759 if (extractLBRStack(TraceIt, Sample->LBRStack)) { 760 if (IgnoreStackSamples) { 761 Sample->CallStack.clear(); 762 } else { 763 // Canonicalize stack leaf to avoid 'random' IP from leaf frame skew LBR 764 // ranges 765 Sample->CallStack.front() = Sample->LBRStack[0].Target; 766 } 767 // Record samples by aggregation 768 AggregatedSamples[Hashable<PerfSample>(Sample)] += Count; 769 } 770 } else { 771 // LBR sample is encoded in single line after stack sample 772 exitWithError("'Hybrid perf sample is corrupted, No LBR sample line"); 773 } 774 } 775 776 void PerfScriptReader::writeUnsymbolizedProfile(StringRef Filename) { 777 std::error_code EC; 778 raw_fd_ostream OS(Filename, EC, llvm::sys::fs::OF_TextWithCRLF); 779 if (EC) 780 exitWithError(EC, Filename); 781 writeUnsymbolizedProfile(OS); 782 } 783 784 // Use ordered map to make the output deterministic 785 using OrderedCounterForPrint = std::map<std::string, SampleCounter *>; 786 787 void PerfScriptReader::writeUnsymbolizedProfile(raw_fd_ostream &OS) { 788 OrderedCounterForPrint OrderedCounters; 789 for (auto &CI : SampleCounters) { 790 OrderedCounters[getContextKeyStr(CI.first.getPtr(), Binary)] = &CI.second; 791 } 792 793 auto SCounterPrinter = [&](RangeSample &Counter, StringRef Separator, 794 uint32_t Indent) { 795 OS.indent(Indent); 796 OS << Counter.size() << "\n"; 797 for (auto &I : Counter) { 798 uint64_t Start = I.first.first; 799 uint64_t End = I.first.second; 800 801 if (!UseOffset || (UseOffset && UseLoadableSegmentAsBase)) { 802 Start = Binary->offsetToVirtualAddr(Start); 803 End = Binary->offsetToVirtualAddr(End); 804 } 805 806 if (UseOffset && UseLoadableSegmentAsBase) { 807 Start -= Binary->getFirstLoadableAddress(); 808 End -= Binary->getFirstLoadableAddress(); 809 } 810 811 OS.indent(Indent); 812 OS << Twine::utohexstr(Start) << Separator << Twine::utohexstr(End) << ":" 813 << I.second << "\n"; 814 } 815 }; 816 817 for (auto &CI : OrderedCounters) { 818 uint32_t Indent = 0; 819 if (ProfileIsCSFlat) { 820 // Context string key 821 OS << "[" << CI.first << "]\n"; 822 Indent = 2; 823 } 824 825 SampleCounter &Counter = *CI.second; 826 SCounterPrinter(Counter.RangeCounter, "-", Indent); 827 SCounterPrinter(Counter.BranchCounter, "->", Indent); 828 } 829 } 830 831 // Format of input: 832 // number of entries in RangeCounter 833 // from_1-to_1:count_1 834 // from_2-to_2:count_2 835 // ...... 836 // from_n-to_n:count_n 837 // number of entries in BranchCounter 838 // src_1->dst_1:count_1 839 // src_2->dst_2:count_2 840 // ...... 841 // src_n->dst_n:count_n 842 void UnsymbolizedProfileReader::readSampleCounters(TraceStream &TraceIt, 843 SampleCounter &SCounters) { 844 auto exitWithErrorForTraceLine = [](TraceStream &TraceIt) { 845 std::string Msg = TraceIt.isAtEoF() 846 ? "Invalid raw profile!" 847 : "Invalid raw profile at line " + 848 Twine(TraceIt.getLineNumber()).str() + ": " + 849 TraceIt.getCurrentLine().str(); 850 exitWithError(Msg); 851 }; 852 auto ReadNumber = [&](uint64_t &Num) { 853 if (TraceIt.isAtEoF()) 854 exitWithErrorForTraceLine(TraceIt); 855 if (TraceIt.getCurrentLine().ltrim().getAsInteger(10, Num)) 856 exitWithErrorForTraceLine(TraceIt); 857 TraceIt.advance(); 858 }; 859 860 auto ReadCounter = [&](RangeSample &Counter, StringRef Separator) { 861 uint64_t Num = 0; 862 ReadNumber(Num); 863 while (Num--) { 864 if (TraceIt.isAtEoF()) 865 exitWithErrorForTraceLine(TraceIt); 866 StringRef Line = TraceIt.getCurrentLine().ltrim(); 867 868 uint64_t Count = 0; 869 auto LineSplit = Line.split(":"); 870 if (LineSplit.second.empty() || LineSplit.second.getAsInteger(10, Count)) 871 exitWithErrorForTraceLine(TraceIt); 872 873 uint64_t Source = 0; 874 uint64_t Target = 0; 875 auto Range = LineSplit.first.split(Separator); 876 if (Range.second.empty() || Range.first.getAsInteger(16, Source) || 877 Range.second.getAsInteger(16, Target)) 878 exitWithErrorForTraceLine(TraceIt); 879 880 if (!UseOffset || (UseOffset && UseLoadableSegmentAsBase)) { 881 uint64_t BaseAddr = 0; 882 if (UseOffset && UseLoadableSegmentAsBase) 883 BaseAddr = Binary->getFirstLoadableAddress(); 884 885 Source = Binary->virtualAddrToOffset(Source + BaseAddr); 886 Target = Binary->virtualAddrToOffset(Target + BaseAddr); 887 } 888 889 Counter[{Source, Target}] += Count; 890 TraceIt.advance(); 891 } 892 }; 893 894 ReadCounter(SCounters.RangeCounter, "-"); 895 ReadCounter(SCounters.BranchCounter, "->"); 896 } 897 898 void UnsymbolizedProfileReader::readUnsymbolizedProfile(StringRef FileName) { 899 TraceStream TraceIt(FileName); 900 while (!TraceIt.isAtEoF()) { 901 std::shared_ptr<StringBasedCtxKey> Key = 902 std::make_shared<StringBasedCtxKey>(); 903 StringRef Line = TraceIt.getCurrentLine(); 904 // Read context stack for CS profile. 905 if (Line.startswith("[")) { 906 ProfileIsCSFlat = true; 907 auto I = ContextStrSet.insert(Line.str()); 908 SampleContext::createCtxVectorFromStr(*I.first, Key->Context); 909 TraceIt.advance(); 910 } 911 auto Ret = 912 SampleCounters.emplace(Hashable<ContextKey>(Key), SampleCounter()); 913 readSampleCounters(TraceIt, Ret.first->second); 914 } 915 } 916 917 void UnsymbolizedProfileReader::parsePerfTraces() { 918 readUnsymbolizedProfile(PerfTraceFile); 919 } 920 921 void PerfScriptReader::computeCounterFromLBR(const PerfSample *Sample, 922 uint64_t Repeat) { 923 SampleCounter &Counter = SampleCounters.begin()->second; 924 uint64_t EndOffeset = 0; 925 for (const LBREntry &LBR : Sample->LBRStack) { 926 assert(LBR.Source != ExternalAddr && 927 "Branch' source should not be an external address, it should be " 928 "converted to aritificial branch."); 929 uint64_t SourceOffset = Binary->virtualAddrToOffset(LBR.Source); 930 uint64_t TargetOffset = LBR.Target == static_cast<uint64_t>(ExternalAddr) 931 ? static_cast<uint64_t>(ExternalAddr) 932 : Binary->virtualAddrToOffset(LBR.Target); 933 934 if (!LBR.IsArtificial && TargetOffset != ExternalAddr) { 935 Counter.recordBranchCount(SourceOffset, TargetOffset, Repeat); 936 } 937 938 // If this not the first LBR, update the range count between TO of current 939 // LBR and FROM of next LBR. 940 uint64_t StartOffset = TargetOffset; 941 if (EndOffeset != 0) { 942 if (StartOffset <= EndOffeset) 943 Counter.recordRangeCount(StartOffset, EndOffeset, Repeat); 944 } 945 EndOffeset = SourceOffset; 946 } 947 } 948 949 void LBRPerfReader::parseSample(TraceStream &TraceIt, uint64_t Count) { 950 std::shared_ptr<PerfSample> Sample = std::make_shared<PerfSample>(); 951 // Parsing LBR stack and populate into PerfSample.LBRStack 952 if (extractLBRStack(TraceIt, Sample->LBRStack)) { 953 warnIfMissingMMap(); 954 // Record LBR only samples by aggregation 955 AggregatedSamples[Hashable<PerfSample>(Sample)] += Count; 956 } 957 } 958 959 void PerfScriptReader::generateUnsymbolizedProfile() { 960 // There is no context for LBR only sample, so initialize one entry with 961 // fake "empty" context key. 962 assert(SampleCounters.empty() && 963 "Sample counter map should be empty before raw profile generation"); 964 std::shared_ptr<StringBasedCtxKey> Key = 965 std::make_shared<StringBasedCtxKey>(); 966 SampleCounters.emplace(Hashable<ContextKey>(Key), SampleCounter()); 967 for (const auto &Item : AggregatedSamples) { 968 const PerfSample *Sample = Item.first.getPtr(); 969 computeCounterFromLBR(Sample, Item.second); 970 } 971 } 972 973 uint64_t PerfScriptReader::parseAggregatedCount(TraceStream &TraceIt) { 974 // The aggregated count is optional, so do not skip the line and return 1 if 975 // it's unmatched 976 uint64_t Count = 1; 977 if (!TraceIt.getCurrentLine().getAsInteger(10, Count)) 978 TraceIt.advance(); 979 return Count; 980 } 981 982 void PerfScriptReader::parseSample(TraceStream &TraceIt) { 983 NumTotalSample++; 984 uint64_t Count = parseAggregatedCount(TraceIt); 985 assert(Count >= 1 && "Aggregated count should be >= 1!"); 986 parseSample(TraceIt, Count); 987 } 988 989 bool PerfScriptReader::extractMMap2EventForBinary(ProfiledBinary *Binary, 990 StringRef Line, 991 MMapEvent &MMap) { 992 // Parse a line like: 993 // PERF_RECORD_MMAP2 2113428/2113428: [0x7fd4efb57000(0x204000) @ 0 994 // 08:04 19532229 3585508847]: r-xp /usr/lib64/libdl-2.17.so 995 constexpr static const char *const Pattern = 996 "PERF_RECORD_MMAP2 ([0-9]+)/[0-9]+: " 997 "\\[(0x[a-f0-9]+)\\((0x[a-f0-9]+)\\) @ " 998 "(0x[a-f0-9]+|0) .*\\]: [-a-z]+ (.*)"; 999 // Field 0 - whole line 1000 // Field 1 - PID 1001 // Field 2 - base address 1002 // Field 3 - mmapped size 1003 // Field 4 - page offset 1004 // Field 5 - binary path 1005 enum EventIndex { 1006 WHOLE_LINE = 0, 1007 PID = 1, 1008 MMAPPED_ADDRESS = 2, 1009 MMAPPED_SIZE = 3, 1010 PAGE_OFFSET = 4, 1011 BINARY_PATH = 5 1012 }; 1013 1014 Regex RegMmap2(Pattern); 1015 SmallVector<StringRef, 6> Fields; 1016 bool R = RegMmap2.match(Line, &Fields); 1017 if (!R) { 1018 std::string ErrorMsg = "Cannot parse mmap event: " + Line.str() + " \n"; 1019 exitWithError(ErrorMsg); 1020 } 1021 Fields[PID].getAsInteger(10, MMap.PID); 1022 Fields[MMAPPED_ADDRESS].getAsInteger(0, MMap.Address); 1023 Fields[MMAPPED_SIZE].getAsInteger(0, MMap.Size); 1024 Fields[PAGE_OFFSET].getAsInteger(0, MMap.Offset); 1025 MMap.BinaryPath = Fields[BINARY_PATH]; 1026 if (ShowMmapEvents) { 1027 outs() << "Mmap: Binary " << MMap.BinaryPath << " loaded at " 1028 << format("0x%" PRIx64 ":", MMap.Address) << " \n"; 1029 } 1030 1031 StringRef BinaryName = llvm::sys::path::filename(MMap.BinaryPath); 1032 return Binary->getName() == BinaryName; 1033 } 1034 1035 void PerfScriptReader::parseMMap2Event(TraceStream &TraceIt) { 1036 MMapEvent MMap; 1037 if (extractMMap2EventForBinary(Binary, TraceIt.getCurrentLine(), MMap)) 1038 updateBinaryAddress(MMap); 1039 TraceIt.advance(); 1040 } 1041 1042 void PerfScriptReader::parseEventOrSample(TraceStream &TraceIt) { 1043 if (isMMap2Event(TraceIt.getCurrentLine())) 1044 parseMMap2Event(TraceIt); 1045 else 1046 parseSample(TraceIt); 1047 } 1048 1049 void PerfScriptReader::parseAndAggregateTrace() { 1050 // Trace line iterator 1051 TraceStream TraceIt(PerfTraceFile); 1052 while (!TraceIt.isAtEoF()) 1053 parseEventOrSample(TraceIt); 1054 } 1055 1056 // A LBR sample is like: 1057 // 40062f 0x5c6313f/0x5c63170/P/-/-/0 0x5c630e7/0x5c63130/P/-/-/0 ... 1058 // A heuristic for fast detection by checking whether a 1059 // leading " 0x" and the '/' exist. 1060 bool PerfScriptReader::isLBRSample(StringRef Line) { 1061 // Skip the leading instruction pointer 1062 SmallVector<StringRef, 32> Records; 1063 Line.trim().split(Records, " ", 2, false); 1064 if (Records.size() < 2) 1065 return false; 1066 if (Records[1].startswith("0x") && Records[1].contains('/')) 1067 return true; 1068 return false; 1069 } 1070 1071 bool PerfScriptReader::isMMap2Event(StringRef Line) { 1072 // Short cut to avoid string find is possible. 1073 if (Line.empty() || Line.size() < 50) 1074 return false; 1075 1076 if (std::isdigit(Line[0])) 1077 return false; 1078 1079 // PERF_RECORD_MMAP2 does not appear at the beginning of the line 1080 // for ` perf script --show-mmap-events -i ...` 1081 return Line.contains("PERF_RECORD_MMAP2"); 1082 } 1083 1084 // The raw hybird sample is like 1085 // e.g. 1086 // 4005dc # call stack leaf 1087 // 400634 1088 // 400684 # call stack root 1089 // 0x4005c8/0x4005dc/P/-/-/0 0x40062f/0x4005b0/P/-/-/0 ... 1090 // ... 0x4005c8/0x4005dc/P/-/-/0 # LBR Entries 1091 // Determine the perfscript contains hybrid samples(call stack + LBRs) by 1092 // checking whether there is a non-empty call stack immediately followed by 1093 // a LBR sample 1094 PerfContent PerfScriptReader::checkPerfScriptType(StringRef FileName) { 1095 TraceStream TraceIt(FileName); 1096 uint64_t FrameAddr = 0; 1097 while (!TraceIt.isAtEoF()) { 1098 // Skip the aggregated count 1099 if (!TraceIt.getCurrentLine().getAsInteger(10, FrameAddr)) 1100 TraceIt.advance(); 1101 1102 // Detect sample with call stack 1103 int32_t Count = 0; 1104 while (!TraceIt.isAtEoF() && 1105 !TraceIt.getCurrentLine().ltrim().getAsInteger(16, FrameAddr)) { 1106 Count++; 1107 TraceIt.advance(); 1108 } 1109 if (!TraceIt.isAtEoF()) { 1110 if (isLBRSample(TraceIt.getCurrentLine())) { 1111 if (Count > 0) 1112 return PerfContent::LBRStack; 1113 else 1114 return PerfContent::LBR; 1115 } 1116 TraceIt.advance(); 1117 } 1118 } 1119 1120 exitWithError("Invalid perf script input!"); 1121 return PerfContent::UnknownContent; 1122 } 1123 1124 void HybridPerfReader::generateUnsymbolizedProfile() { 1125 ProfileIsCSFlat = !IgnoreStackSamples; 1126 if (ProfileIsCSFlat) 1127 unwindSamples(); 1128 else 1129 PerfScriptReader::generateUnsymbolizedProfile(); 1130 } 1131 1132 void PerfScriptReader::warnTruncatedStack() { 1133 if (ShowDetailedWarning) { 1134 for (auto Address : InvalidReturnAddresses) { 1135 WithColor::warning() 1136 << "Truncated stack sample due to invalid return address at " 1137 << format("0x%" PRIx64, Address) 1138 << ", likely caused by frame pointer omission\n"; 1139 } 1140 } 1141 emitWarningSummary( 1142 InvalidReturnAddresses.size(), AggregatedSamples.size(), 1143 "of truncated stack samples due to invalid return address, " 1144 "likely caused by frame pointer omission."); 1145 } 1146 1147 void PerfScriptReader::warnInvalidRange() { 1148 std::unordered_map<std::pair<uint64_t, uint64_t>, uint64_t, 1149 pair_hash<uint64_t, uint64_t>> 1150 Ranges; 1151 1152 for (const auto &Item : AggregatedSamples) { 1153 const PerfSample *Sample = Item.first.getPtr(); 1154 uint64_t Count = Item.second; 1155 uint64_t EndOffeset = 0; 1156 for (const LBREntry &LBR : Sample->LBRStack) { 1157 uint64_t SourceOffset = Binary->virtualAddrToOffset(LBR.Source); 1158 uint64_t StartOffset = Binary->virtualAddrToOffset(LBR.Target); 1159 if (EndOffeset != 0) 1160 Ranges[{StartOffset, EndOffeset}] += Count; 1161 EndOffeset = SourceOffset; 1162 } 1163 } 1164 1165 if (Ranges.empty()) { 1166 WithColor::warning() << "No samples in perf script!\n"; 1167 return; 1168 } 1169 1170 auto WarnInvalidRange = 1171 [&](uint64_t StartOffset, uint64_t EndOffset, StringRef Msg) { 1172 if (!ShowDetailedWarning) 1173 return; 1174 WithColor::warning() 1175 << "[" 1176 << format("%8" PRIx64, Binary->offsetToVirtualAddr(StartOffset)) 1177 << "," 1178 << format("%8" PRIx64, Binary->offsetToVirtualAddr(EndOffset)) 1179 << "]: " << Msg << "\n"; 1180 }; 1181 1182 const char *EndNotBoundaryMsg = "Range is not on instruction boundary, " 1183 "likely due to profile and binary mismatch."; 1184 const char *DanglingRangeMsg = "Range does not belong to any functions, " 1185 "likely from PLT, .init or .fini section."; 1186 const char *RangeCrossFuncMsg = 1187 "Fall through range should not cross function boundaries, likely due to " 1188 "profile and binary mismatch."; 1189 const char *BogusRangeMsg = "Range start is after range end."; 1190 1191 uint64_t TotalRangeNum = 0; 1192 uint64_t InstNotBoundary = 0; 1193 uint64_t UnmatchedRange = 0; 1194 uint64_t RangeCrossFunc = 0; 1195 uint64_t BogusRange = 0; 1196 1197 for (auto &I : Ranges) { 1198 uint64_t StartOffset = I.first.first; 1199 uint64_t EndOffset = I.first.second; 1200 TotalRangeNum += I.second; 1201 1202 if (!Binary->offsetIsCode(StartOffset) || 1203 !Binary->offsetIsTransfer(EndOffset)) { 1204 InstNotBoundary += I.second; 1205 WarnInvalidRange(StartOffset, EndOffset, EndNotBoundaryMsg); 1206 } 1207 1208 auto *FRange = Binary->findFuncRangeForOffset(StartOffset); 1209 if (!FRange) { 1210 UnmatchedRange += I.second; 1211 WarnInvalidRange(StartOffset, EndOffset, DanglingRangeMsg); 1212 continue; 1213 } 1214 1215 if (EndOffset >= FRange->EndOffset) { 1216 RangeCrossFunc += I.second; 1217 WarnInvalidRange(StartOffset, EndOffset, RangeCrossFuncMsg); 1218 } 1219 1220 if (StartOffset > EndOffset) { 1221 BogusRange += I.second; 1222 WarnInvalidRange(StartOffset, EndOffset, BogusRangeMsg); 1223 } 1224 } 1225 1226 emitWarningSummary( 1227 InstNotBoundary, TotalRangeNum, 1228 "of samples are from ranges that are not on instruction boundary."); 1229 emitWarningSummary( 1230 UnmatchedRange, TotalRangeNum, 1231 "of samples are from ranges that do not belong to any functions."); 1232 emitWarningSummary( 1233 RangeCrossFunc, TotalRangeNum, 1234 "of samples are from ranges that do cross function boundaries."); 1235 emitWarningSummary( 1236 BogusRange, TotalRangeNum, 1237 "of samples are from ranges that have range start after range end."); 1238 } 1239 1240 void PerfScriptReader::parsePerfTraces() { 1241 // Parse perf traces and do aggregation. 1242 parseAndAggregateTrace(); 1243 1244 emitWarningSummary(NumLeafExternalFrame, NumTotalSample, 1245 "of samples have leaf external frame in call stack."); 1246 emitWarningSummary(NumLeadingOutgoingLBR, NumTotalSample, 1247 "of samples have leading external LBR."); 1248 1249 // Generate unsymbolized profile. 1250 warnTruncatedStack(); 1251 warnInvalidRange(); 1252 generateUnsymbolizedProfile(); 1253 AggregatedSamples.clear(); 1254 1255 if (SkipSymbolization) 1256 writeUnsymbolizedProfile(OutputFilename); 1257 } 1258 1259 } // end namespace sampleprof 1260 } // end namespace llvm 1261