| 1 | //===-- ProfiledBinary.h - Binary decoder -----------------------*- 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_PROFILEDBINARY_H |
| 10 | #define LLVM_TOOLS_LLVM_PROFGEN_PROFILEDBINARY_H |
| 11 | |
| 12 | #include "CallContext.h" |
| 13 | #include "ErrorHandling.h" |
| 14 | #include "llvm/ADT/DenseMap.h" |
| 15 | #include "llvm/ADT/StringRef.h" |
| 16 | #include "llvm/ADT/StringSet.h" |
| 17 | #include "llvm/DebugInfo/DWARF/DWARFContext.h" |
| 18 | #include "llvm/DebugInfo/Symbolize/Symbolize.h" |
| 19 | #include "llvm/MC/MCAsmInfo.h" |
| 20 | #include "llvm/MC/MCContext.h" |
| 21 | #include "llvm/MC/MCDisassembler/MCDisassembler.h" |
| 22 | #include "llvm/MC/MCInst.h" |
| 23 | #include "llvm/MC/MCInstPrinter.h" |
| 24 | #include "llvm/MC/MCInstrAnalysis.h" |
| 25 | #include "llvm/MC/MCInstrInfo.h" |
| 26 | #include "llvm/MC/MCObjectFileInfo.h" |
| 27 | #include "llvm/MC/MCPseudoProbe.h" |
| 28 | #include "llvm/MC/MCRegisterInfo.h" |
| 29 | #include "llvm/MC/MCSubtargetInfo.h" |
| 30 | #include "llvm/MC/MCTargetOptions.h" |
| 31 | #include "llvm/Object/ELFObjectFile.h" |
| 32 | #include "llvm/ProfileData/SampleProf.h" |
| 33 | #include "llvm/Support/CommandLine.h" |
| 34 | #include "llvm/Support/Path.h" |
| 35 | #include "llvm/Transforms/IPO/SampleContextTracker.h" |
| 36 | #include <map> |
| 37 | #include <set> |
| 38 | #include <sstream> |
| 39 | #include <string> |
| 40 | #include <unordered_map> |
| 41 | #include <unordered_set> |
| 42 | #include <vector> |
| 43 | |
| 44 | namespace llvm { |
| 45 | namespace sampleprof { |
| 46 | |
| 47 | class ProfiledBinary; |
| 48 | class MissingFrameInferrer; |
| 49 | |
| 50 | struct InstructionPointer { |
| 51 | const ProfiledBinary *Binary; |
| 52 | // Address of the executable segment of the binary. |
| 53 | uint64_t Address; |
| 54 | // Index to the sorted code address array of the binary. |
| 55 | uint64_t Index = 0; |
| 56 | InstructionPointer(const ProfiledBinary *Binary, uint64_t Address, |
| 57 | bool RoundToNext = false); |
| 58 | bool advance(); |
| 59 | bool backward(); |
| 60 | void update(uint64_t Addr); |
| 61 | }; |
| 62 | |
| 63 | // The special frame addresses. |
| 64 | enum SpecialFrameAddr { |
| 65 | // Dummy root of frame trie. |
| 66 | DummyRoot = 0, |
| 67 | // Represent all the addresses outside of current binary. |
| 68 | // This's also used to indicate the call stack should be truncated since this |
| 69 | // isn't a real call context the compiler will see. |
| 70 | ExternalAddr = 1, |
| 71 | }; |
| 72 | |
| 73 | using RangesTy = std::vector<std::pair<uint64_t, uint64_t>>; |
| 74 | |
| 75 | enum DwarfNameStatus { |
| 76 | // Dwarf name matches with the symbol table (or symbol table just doesn't have |
| 77 | // this entry) |
| 78 | Matched = 0, |
| 79 | // Dwarf name is missing, but we fixed it with the name from symbol table |
| 80 | Missing = 1, |
| 81 | // Symbol table has different names on this. Log these GUIDs in |
| 82 | // AlternativeFunctionGUIDs |
| 83 | Mismatch = 2, |
| 84 | }; |
| 85 | |
| 86 | struct BinaryFunction { |
| 87 | StringRef FuncName; |
| 88 | // End of range is an exclusive bound. |
| 89 | RangesTy Ranges; |
| 90 | DwarfNameStatus NameStatus = DwarfNameStatus::Matched; |
| 91 | |
| 92 | uint64_t getFuncSize() { |
| 93 | uint64_t Sum = 0; |
| 94 | for (auto &R : Ranges) { |
| 95 | Sum += R.second - R.first; |
| 96 | } |
| 97 | return Sum; |
| 98 | } |
| 99 | }; |
| 100 | |
| 101 | // Info about function range. A function can be split into multiple |
| 102 | // non-continuous ranges, each range corresponds to one FuncRange. |
| 103 | struct FuncRange { |
| 104 | uint64_t StartAddress; |
| 105 | // EndAddress is an exclusive bound. |
| 106 | uint64_t EndAddress; |
| 107 | // Function the range belongs to |
| 108 | BinaryFunction *Func; |
| 109 | // Whether the start address is the real entry of the function. |
| 110 | bool IsFuncEntry = false; |
| 111 | |
| 112 | StringRef getFuncName() { return Func->FuncName; } |
| 113 | }; |
| 114 | |
| 115 | // PrologEpilog address tracker, used to filter out broken stack samples |
| 116 | // Currently we use a heuristic size (two) to infer prolog and epilog |
| 117 | // based on the start address and return address. In the future, |
| 118 | // we will switch to Dwarf CFI based tracker |
| 119 | struct PrologEpilogTracker { |
| 120 | // A set of prolog and epilog addresses. Used by virtual unwinding. |
| 121 | std::unordered_set<uint64_t> PrologEpilogSet; |
| 122 | ProfiledBinary *Binary; |
| 123 | PrologEpilogTracker(ProfiledBinary *Bin) : Binary(Bin){}; |
| 124 | |
| 125 | // Take the two addresses from the start of function as prolog |
| 126 | void |
| 127 | inferPrologAddresses(std::map<uint64_t, FuncRange> &FuncStartAddressMap) { |
| 128 | for (auto I : FuncStartAddressMap) { |
| 129 | PrologEpilogSet.insert(x: I.first); |
| 130 | InstructionPointer IP(Binary, I.first); |
| 131 | if (!IP.advance()) |
| 132 | break; |
| 133 | PrologEpilogSet.insert(x: IP.Address); |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | // Take the last two addresses before the return address as epilog |
| 138 | void inferEpilogAddresses(std::unordered_set<uint64_t> &RetAddrs) { |
| 139 | for (auto Addr : RetAddrs) { |
| 140 | PrologEpilogSet.insert(x: Addr); |
| 141 | InstructionPointer IP(Binary, Addr); |
| 142 | if (!IP.backward()) |
| 143 | break; |
| 144 | PrologEpilogSet.insert(x: IP.Address); |
| 145 | } |
| 146 | } |
| 147 | }; |
| 148 | |
| 149 | // Track function byte size under different context (outlined version as well as |
| 150 | // various inlined versions). It also provides query support to get function |
| 151 | // size with the best matching context, which is used to help pre-inliner use |
| 152 | // accurate post-optimization size to make decisions. |
| 153 | // TODO: If an inlinee is completely optimized away, ideally we should have zero |
| 154 | // for its context size, currently we would misss such context since it doesn't |
| 155 | // have instructions. To fix this, we need to mark all inlinee with entry probe |
| 156 | // but without instructions as having zero size. |
| 157 | class BinarySizeContextTracker { |
| 158 | public: |
| 159 | // Add instruction with given size to a context |
| 160 | void addInstructionForContext(const SampleContextFrameVector &Context, |
| 161 | uint32_t InstrSize); |
| 162 | |
| 163 | // Get function size with a specific context. When there's no exact match |
| 164 | // for the given context, try to retrieve the size of that function from |
| 165 | // closest matching context. |
| 166 | uint32_t getFuncSizeForContext(const ContextTrieNode *Context); |
| 167 | |
| 168 | // For inlinees that are full optimized away, we can establish zero size using |
| 169 | // their remaining probes. |
| 170 | void trackInlineesOptimizedAway(MCPseudoProbeDecoder &ProbeDecoder); |
| 171 | |
| 172 | using ProbeFrameStack = SmallVector<std::pair<StringRef, uint32_t>>; |
| 173 | void |
| 174 | trackInlineesOptimizedAway(MCPseudoProbeDecoder &ProbeDecoder, |
| 175 | const MCDecodedPseudoProbeInlineTree &ProbeNode, |
| 176 | ProbeFrameStack &Context); |
| 177 | |
| 178 | void dump() { RootContext.dumpTree(); } |
| 179 | |
| 180 | private: |
| 181 | // Root node for context trie tree, node that this is a reverse context trie |
| 182 | // with callee as parent and caller as child. This way we can traverse from |
| 183 | // root to find the best/longest matching context if an exact match does not |
| 184 | // exist. It gives us the best possible estimate for function's post-inline, |
| 185 | // post-optimization byte size. |
| 186 | ContextTrieNode RootContext; |
| 187 | }; |
| 188 | |
| 189 | using AddressRange = std::pair<uint64_t, uint64_t>; |
| 190 | |
| 191 | // The parsed MMap event |
| 192 | struct MMapEvent { |
| 193 | int64_t PID = 0; |
| 194 | uint64_t Address = 0; |
| 195 | uint64_t Size = 0; |
| 196 | uint64_t Offset = 0; |
| 197 | StringRef MemProtectionFlag; |
| 198 | StringRef BinaryPath; |
| 199 | }; |
| 200 | |
| 201 | class ProfiledBinary { |
| 202 | // Absolute path of the executable binary. |
| 203 | std::string Path; |
| 204 | // Path of the debug info binary. |
| 205 | std::string DebugBinaryPath; |
| 206 | // Path of the pseudo probe binary, either Path or DebugBinaryPath if present. |
| 207 | StringRef PseudoProbeBinPath; |
| 208 | // The target triple. |
| 209 | Triple TheTriple; |
| 210 | // Path of symbolizer path which should be pointed to binary with debug info. |
| 211 | StringRef SymbolizerPath; |
| 212 | // Options used to configure the symbolizer |
| 213 | symbolize::LLVMSymbolizer::Options SymbolizerOpts; |
| 214 | // The runtime base address that the first executable segment is loaded at. |
| 215 | uint64_t BaseAddress = 0; |
| 216 | // The runtime base address that the first loadabe segment is loaded at. |
| 217 | uint64_t FirstLoadableAddress = 0; |
| 218 | // The preferred load address of each executable segment. |
| 219 | std::vector<uint64_t> PreferredTextSegmentAddresses; |
| 220 | // The file offset of each executable segment. |
| 221 | std::vector<uint64_t> TextSegmentOffsets; |
| 222 | |
| 223 | // Mutiple MC component info |
| 224 | std::unique_ptr<const MCRegisterInfo> MRI; |
| 225 | std::unique_ptr<const MCAsmInfo> AsmInfo; |
| 226 | std::unique_ptr<const MCSubtargetInfo> STI; |
| 227 | std::unique_ptr<const MCInstrInfo> MII; |
| 228 | std::unique_ptr<MCDisassembler> DisAsm; |
| 229 | std::unique_ptr<const MCInstrAnalysis> MIA; |
| 230 | std::unique_ptr<MCInstPrinter> IPrinter; |
| 231 | // A list of text sections sorted by start RVA and size. Used to check |
| 232 | // if a given RVA is a valid code address. |
| 233 | std::set<std::pair<uint64_t, uint64_t>> TextSections; |
| 234 | |
| 235 | // A map of mapping function name to BinaryFunction info. |
| 236 | std::unordered_map<std::string, BinaryFunction> BinaryFunctions; |
| 237 | |
| 238 | // Lookup BinaryFunctions using the function name's MD5 hash. Needed if the |
| 239 | // profile is using MD5. |
| 240 | std::unordered_map<uint64_t, BinaryFunction *> HashBinaryFunctions; |
| 241 | |
| 242 | // A list of binary functions that have samples. |
| 243 | std::unordered_set<const BinaryFunction *> ProfiledFunctions; |
| 244 | |
| 245 | // GUID to symbol start address map |
| 246 | DenseMap<uint64_t, uint64_t> SymbolStartAddrs; |
| 247 | |
| 248 | // Binary function to GUID mapping that stores the alternative names in symbol |
| 249 | // table, despite the original name from DWARF info |
| 250 | std::unordered_multimap<const BinaryFunction *, uint64_t> |
| 251 | AlternativeFunctionGUIDs; |
| 252 | |
| 253 | // Mapping of profiled binary function to its pseudo probe name |
| 254 | std::unordered_map<const BinaryFunction *, StringRef> PseudoProbeNames; |
| 255 | |
| 256 | // These maps are for temporary use of warning diagnosis. |
| 257 | DenseSet<int64_t> AddrsWithMultipleSymbols; |
| 258 | DenseSet<std::pair<uint64_t, uint64_t>> AddrsWithInvalidInstruction; |
| 259 | |
| 260 | // Start address to symbol GUID map |
| 261 | std::unordered_multimap<uint64_t, uint64_t> StartAddrToSymMap; |
| 262 | |
| 263 | // An ordered map of mapping function's start address to function range |
| 264 | // relevant info. Currently to determine if the offset of ELF/COFF is the |
| 265 | // start of a real function, we leverage the function range info from DWARF. |
| 266 | std::map<uint64_t, FuncRange> StartAddrToFuncRangeMap; |
| 267 | |
| 268 | // Address to context location map. Used to expand the context. |
| 269 | std::unordered_map<uint64_t, SampleContextFrameVector> AddressToLocStackMap; |
| 270 | |
| 271 | // Address to instruction size map. Also used for quick Address lookup. |
| 272 | std::unordered_map<uint64_t, uint64_t> AddressToInstSizeMap; |
| 273 | |
| 274 | // An array of Addresses of all instructions sorted in increasing order. The |
| 275 | // sorting is needed to fast advance to the next forward/backward instruction. |
| 276 | std::vector<uint64_t> CodeAddressVec; |
| 277 | // A set of call instruction addresses. Used by virtual unwinding. |
| 278 | std::unordered_set<uint64_t> CallAddressSet; |
| 279 | // A set of return instruction addresses. Used by virtual unwinding. |
| 280 | std::unordered_set<uint64_t> RetAddressSet; |
| 281 | // An ordered set of unconditional branch instruction addresses. |
| 282 | std::set<uint64_t> UncondBranchAddrSet; |
| 283 | // A set of branch instruction addresses. |
| 284 | std::unordered_set<uint64_t> BranchAddressSet; |
| 285 | |
| 286 | // Estimate and track function prolog and epilog ranges. |
| 287 | PrologEpilogTracker ProEpilogTracker; |
| 288 | |
| 289 | // Infer missing frames due to compiler optimizations such as tail call |
| 290 | // elimination. |
| 291 | std::unique_ptr<MissingFrameInferrer> MissingContextInferrer; |
| 292 | |
| 293 | // Track function sizes under different context |
| 294 | BinarySizeContextTracker FuncSizeTracker; |
| 295 | |
| 296 | // The symbolizer used to get inline context for an instruction. |
| 297 | std::unique_ptr<symbolize::LLVMSymbolizer> Symbolizer; |
| 298 | |
| 299 | // String table owning function name strings created from the symbolizer. |
| 300 | std::unordered_set<std::string> NameStrings; |
| 301 | |
| 302 | // MMap events for PT_LOAD segments without 'x' memory protection flag. |
| 303 | std::map<uint64_t, MMapEvent, std::greater<uint64_t>> NonTextMMapEvents; |
| 304 | |
| 305 | // Records the file offset, file size and virtual address of program headers. |
| 306 | struct PhdrInfo { |
| 307 | uint64_t FileOffset; |
| 308 | uint64_t FileSz; |
| 309 | uint64_t VirtualAddr; |
| 310 | }; |
| 311 | |
| 312 | // Program header information for non-text PT_LOAD segments. |
| 313 | SmallVector<PhdrInfo> NonTextPhdrInfo; |
| 314 | |
| 315 | // A collection of functions to print disassembly for. |
| 316 | StringSet<> DisassembleFunctionSet; |
| 317 | |
| 318 | // Pseudo probe decoder |
| 319 | MCPseudoProbeDecoder ProbeDecoder; |
| 320 | |
| 321 | // Function name to probe frame map for top-level outlined functions. |
| 322 | StringMap<MCDecodedPseudoProbeInlineTree *> TopLevelProbeFrameMap; |
| 323 | |
| 324 | bool UseFSDiscriminator = false; |
| 325 | |
| 326 | // Whether we need to symbolize all instructions to get function context size. |
| 327 | bool TrackFuncContextSize = false; |
| 328 | |
| 329 | // Whether this is a kernel image; |
| 330 | bool IsKernel = false; |
| 331 | |
| 332 | // Indicate if the base loading address is parsed from the mmap event or uses |
| 333 | // the preferred address |
| 334 | bool IsLoadedByMMap = false; |
| 335 | // Use to avoid redundant warning. |
| 336 | bool MissingMMapWarned = false; |
| 337 | |
| 338 | bool IsCOFF = false; |
| 339 | |
| 340 | void setPreferredTextSegmentAddresses(const object::ObjectFile *O); |
| 341 | |
| 342 | // LLVMSymbolizer's symbolize{Code, Data} interfaces requires a section index |
| 343 | // for each address to be symbolized. This is a helper function to |
| 344 | // construct a SectionedAddress object with the given address and section |
| 345 | // index. The section index is set to UndefSection by default. |
| 346 | static object::SectionedAddress getSectionedAddress( |
| 347 | uint64_t Address, |
| 348 | uint64_t SectionIndex = object::SectionedAddress::UndefSection) { |
| 349 | return object::SectionedAddress{.Address: Address, .SectionIndex: SectionIndex}; |
| 350 | } |
| 351 | |
| 352 | template <class ELFT> |
| 353 | void setPreferredTextSegmentAddresses(const object::ELFFile<ELFT> &Obj, |
| 354 | StringRef FileName); |
| 355 | void setPreferredTextSegmentAddresses(const object::COFFObjectFile *Obj, |
| 356 | StringRef FileName); |
| 357 | |
| 358 | // Return true if pseudo probe in Obj is usable. |
| 359 | bool checkPseudoProbe(const object::ObjectFile *Obj, StringRef ObjPath); |
| 360 | |
| 361 | void decodePseudoProbe(const object::ObjectFile *Obj); |
| 362 | |
| 363 | void checkUseFSDiscriminator( |
| 364 | const object::ObjectFile *Obj, |
| 365 | std::map<object::SectionRef, SectionSymbolsTy> &AllSymbols); |
| 366 | |
| 367 | // Set up disassembler and related components. |
| 368 | void setUpDisassembler(const object::ObjectFile *Obj); |
| 369 | symbolize::LLVMSymbolizer::Options getSymbolizerOpts() const; |
| 370 | |
| 371 | // Load debug info of subprograms from DWARF section. |
| 372 | void loadSymbolsFromDWARF(object::ObjectFile &Obj); |
| 373 | |
| 374 | // Load debug info from DWARF unit. |
| 375 | void loadSymbolsFromDWARFUnit(DWARFUnit &CompilationUnit); |
| 376 | |
| 377 | // Create symbol to its start address mapping. |
| 378 | void populateSymbolAddressList(const object::ObjectFile *O); |
| 379 | |
| 380 | // Load functions from its symbol table (when DWARF info is missing). |
| 381 | void loadSymbolsFromSymtab(const object::ObjectFile *O); |
| 382 | |
| 383 | // A function may be spilt into multiple non-continuous address ranges. We use |
| 384 | // this to set whether start a function range is the real entry of the |
| 385 | // function and also set false to the non-function label. |
| 386 | void setIsFuncEntry(FuncRange *FRange, StringRef RangeSymName); |
| 387 | |
| 388 | // Warn if no entry range exists in the function. |
| 389 | void warnNoFuncEntry(); |
| 390 | |
| 391 | /// Dissassemble the text section and build various address maps. |
| 392 | void disassemble(const object::ObjectFile *O); |
| 393 | |
| 394 | /// Helper function to dissassemble the symbol and extract info for unwinding |
| 395 | bool dissassembleSymbol(std::size_t SI, ArrayRef<uint8_t> Bytes, |
| 396 | SectionSymbolsTy &Symbols, |
| 397 | const object::SectionRef &Section); |
| 398 | /// Symbolize a given instruction pointer and return a full call context. |
| 399 | SampleContextFrameVector symbolize(const InstructionPointer &IP, |
| 400 | bool UseCanonicalFnName = false, |
| 401 | bool UseProbeDiscriminator = false); |
| 402 | /// Decode the interesting parts of the binary and build internal data |
| 403 | /// structures. On high level, the parts of interest are: |
| 404 | /// 1. Text sections, including the main code section and the PLT |
| 405 | /// entries that will be used to handle cross-module call transitions. |
| 406 | /// 2. The .debug_line section, used by Dwarf-based profile generation. |
| 407 | /// 3. Pseudo probe related sections, used by probe-based profile |
| 408 | /// generation. |
| 409 | void load(); |
| 410 | |
| 411 | public: |
| 412 | ProfiledBinary(const StringRef ExeBinPath, const StringRef DebugBinPath); |
| 413 | ~ProfiledBinary(); |
| 414 | |
| 415 | /// Symbolize an address and return the symbol name. The returned StringRef is |
| 416 | /// owned by this ProfiledBinary object. |
| 417 | StringRef symbolizeDataAddress(uint64_t Address); |
| 418 | |
| 419 | void decodePseudoProbe(); |
| 420 | |
| 421 | StringRef getPath() const { return Path; } |
| 422 | StringRef getName() const { return llvm::sys::path::filename(path: Path); } |
| 423 | uint64_t getBaseAddress() const { return BaseAddress; } |
| 424 | void setBaseAddress(uint64_t Address) { BaseAddress = Address; } |
| 425 | |
| 426 | bool isCOFF() const { return IsCOFF; } |
| 427 | |
| 428 | // Canonicalize to use preferred load address as base address. |
| 429 | uint64_t canonicalizeVirtualAddress(uint64_t Address) { |
| 430 | return Address - BaseAddress + getPreferredBaseAddress(); |
| 431 | } |
| 432 | // Return the preferred load address for the first executable segment. |
| 433 | uint64_t getPreferredBaseAddress() const { |
| 434 | return PreferredTextSegmentAddresses[0]; |
| 435 | } |
| 436 | // Return the preferred load address for the first loadable segment. |
| 437 | uint64_t getFirstLoadableAddress() const { return FirstLoadableAddress; } |
| 438 | // Return the file offset for the first executable segment. |
| 439 | uint64_t getTextSegmentOffset() const { return TextSegmentOffsets[0]; } |
| 440 | const std::vector<uint64_t> &getPreferredTextSegmentAddresses() const { |
| 441 | return PreferredTextSegmentAddresses; |
| 442 | } |
| 443 | const std::vector<uint64_t> &getTextSegmentOffsets() const { |
| 444 | return TextSegmentOffsets; |
| 445 | } |
| 446 | |
| 447 | uint64_t getInstSize(uint64_t Address) const { |
| 448 | auto I = AddressToInstSizeMap.find(x: Address); |
| 449 | if (I == AddressToInstSizeMap.end()) |
| 450 | return 0; |
| 451 | return I->second; |
| 452 | } |
| 453 | |
| 454 | bool addressIsCode(uint64_t Address) const { |
| 455 | return AddressToInstSizeMap.find(x: Address) != AddressToInstSizeMap.end(); |
| 456 | } |
| 457 | |
| 458 | bool addressIsCall(uint64_t Address) const { |
| 459 | return CallAddressSet.count(x: Address); |
| 460 | } |
| 461 | bool addressIsReturn(uint64_t Address) const { |
| 462 | return RetAddressSet.count(x: Address); |
| 463 | } |
| 464 | bool addressInPrologEpilog(uint64_t Address) const { |
| 465 | return ProEpilogTracker.PrologEpilogSet.count(x: Address); |
| 466 | } |
| 467 | |
| 468 | bool addressIsTransfer(uint64_t Address) { |
| 469 | return BranchAddressSet.count(x: Address) || RetAddressSet.count(x: Address) || |
| 470 | CallAddressSet.count(x: Address); |
| 471 | } |
| 472 | |
| 473 | bool rangeCrossUncondBranch(uint64_t Start, uint64_t End) { |
| 474 | if (Start >= End) |
| 475 | return false; |
| 476 | auto R = UncondBranchAddrSet.lower_bound(x: Start); |
| 477 | return R != UncondBranchAddrSet.end() && *R < End; |
| 478 | } |
| 479 | |
| 480 | uint64_t getAddressforIndex(uint64_t Index) const { |
| 481 | return CodeAddressVec[Index]; |
| 482 | } |
| 483 | |
| 484 | size_t getCodeAddrVecSize() const { return CodeAddressVec.size(); } |
| 485 | |
| 486 | bool usePseudoProbes() const { return !PseudoProbeBinPath.empty(); } |
| 487 | bool useFSDiscriminator() const { return UseFSDiscriminator; } |
| 488 | bool isKernel() const { return IsKernel; } |
| 489 | |
| 490 | static bool isKernelImageName(StringRef BinaryName) { |
| 491 | return BinaryName == "[kernel.kallsyms]" || |
| 492 | BinaryName == "[kernel.kallsyms]_stext" || |
| 493 | BinaryName == "[kernel.kallsyms]_text" ; |
| 494 | } |
| 495 | |
| 496 | // Get the index in CodeAddressVec for the address |
| 497 | // As we might get an address which is not the code |
| 498 | // here it would round to the next valid code address by |
| 499 | // using lower bound operation |
| 500 | uint32_t getIndexForAddr(uint64_t Address) const { |
| 501 | auto Low = llvm::lower_bound(Range: CodeAddressVec, Value&: Address); |
| 502 | return Low - CodeAddressVec.begin(); |
| 503 | } |
| 504 | |
| 505 | uint64_t getCallAddrFromFrameAddr(uint64_t FrameAddr) const { |
| 506 | if (FrameAddr == ExternalAddr) |
| 507 | return ExternalAddr; |
| 508 | auto I = getIndexForAddr(Address: FrameAddr); |
| 509 | FrameAddr = I ? getAddressforIndex(Index: I - 1) : 0; |
| 510 | if (FrameAddr && addressIsCall(Address: FrameAddr)) |
| 511 | return FrameAddr; |
| 512 | return 0; |
| 513 | } |
| 514 | |
| 515 | FuncRange *findFuncRangeForStartAddr(uint64_t Address) { |
| 516 | auto I = StartAddrToFuncRangeMap.find(x: Address); |
| 517 | if (I == StartAddrToFuncRangeMap.end()) |
| 518 | return nullptr; |
| 519 | return &I->second; |
| 520 | } |
| 521 | |
| 522 | // Binary search the function range which includes the input address. |
| 523 | FuncRange *findFuncRange(uint64_t Address) { |
| 524 | auto I = StartAddrToFuncRangeMap.upper_bound(x: Address); |
| 525 | if (I == StartAddrToFuncRangeMap.begin()) |
| 526 | return nullptr; |
| 527 | I--; |
| 528 | |
| 529 | if (Address >= I->second.EndAddress) |
| 530 | return nullptr; |
| 531 | |
| 532 | return &I->second; |
| 533 | } |
| 534 | |
| 535 | // Get all ranges of one function. |
| 536 | RangesTy getRanges(uint64_t Address) { |
| 537 | auto *FRange = findFuncRange(Address); |
| 538 | // Ignore the range which falls into plt section or system lib. |
| 539 | if (!FRange) |
| 540 | return RangesTy(); |
| 541 | |
| 542 | return FRange->Func->Ranges; |
| 543 | } |
| 544 | |
| 545 | const std::unordered_map<std::string, BinaryFunction> & |
| 546 | getAllBinaryFunctions() { |
| 547 | return BinaryFunctions; |
| 548 | } |
| 549 | |
| 550 | std::unordered_set<const BinaryFunction *> &getProfiledFunctions() { |
| 551 | return ProfiledFunctions; |
| 552 | } |
| 553 | |
| 554 | void setProfiledFunctions(std::unordered_set<const BinaryFunction *> &Funcs) { |
| 555 | ProfiledFunctions = Funcs; |
| 556 | } |
| 557 | |
| 558 | BinaryFunction *getBinaryFunction(FunctionId FName) { |
| 559 | if (FName.isStringRef()) { |
| 560 | auto I = BinaryFunctions.find(x: FName.str()); |
| 561 | if (I == BinaryFunctions.end()) |
| 562 | return nullptr; |
| 563 | return &I->second; |
| 564 | } |
| 565 | auto I = HashBinaryFunctions.find(x: FName.getHashCode()); |
| 566 | if (I == HashBinaryFunctions.end()) |
| 567 | return nullptr; |
| 568 | return I->second; |
| 569 | } |
| 570 | |
| 571 | uint32_t getFuncSizeForContext(const ContextTrieNode *ContextNode) { |
| 572 | return FuncSizeTracker.getFuncSizeForContext(Context: ContextNode); |
| 573 | } |
| 574 | |
| 575 | void inferMissingFrames(const SmallVectorImpl<uint64_t> &Context, |
| 576 | SmallVectorImpl<uint64_t> &NewContext); |
| 577 | |
| 578 | // Load the symbols from debug table and populate into symbol list. |
| 579 | void populateSymbolListFromDWARF(ProfileSymbolList &SymbolList); |
| 580 | |
| 581 | SampleContextFrameVector |
| 582 | getFrameLocationStack(uint64_t Address, bool UseProbeDiscriminator = false) { |
| 583 | InstructionPointer IP(this, Address); |
| 584 | return symbolize(IP, UseCanonicalFnName: SymbolizerOpts.UseSymbolTable, UseProbeDiscriminator); |
| 585 | } |
| 586 | |
| 587 | const SampleContextFrameVector & |
| 588 | getCachedFrameLocationStack(uint64_t Address, |
| 589 | bool UseProbeDiscriminator = false) { |
| 590 | auto I = AddressToLocStackMap.emplace(args&: Address, args: SampleContextFrameVector()); |
| 591 | if (I.second) { |
| 592 | I.first->second = getFrameLocationStack(Address, UseProbeDiscriminator); |
| 593 | } |
| 594 | return I.first->second; |
| 595 | } |
| 596 | |
| 597 | std::optional<SampleContextFrame> getInlineLeafFrameLoc(uint64_t Address) { |
| 598 | const auto &Stack = getCachedFrameLocationStack(Address); |
| 599 | if (Stack.empty()) |
| 600 | return {}; |
| 601 | return Stack.back(); |
| 602 | } |
| 603 | |
| 604 | void flushSymbolizer() { Symbolizer.reset(); } |
| 605 | |
| 606 | MissingFrameInferrer *getMissingContextInferrer() { |
| 607 | return MissingContextInferrer.get(); |
| 608 | } |
| 609 | |
| 610 | // Compare two addresses' inline context |
| 611 | bool inlineContextEqual(uint64_t Add1, uint64_t Add2); |
| 612 | |
| 613 | // Get the full context of the current stack with inline context filled in. |
| 614 | // It will search the disassembling info stored in AddressToLocStackMap. This |
| 615 | // is used as the key of function sample map |
| 616 | SampleContextFrameVector |
| 617 | getExpandedContext(const SmallVectorImpl<uint64_t> &Stack, |
| 618 | bool &WasLeafInlined); |
| 619 | // Go through instructions among the given range and record its size for the |
| 620 | // inline context. |
| 621 | void computeInlinedContextSizeForRange(uint64_t StartAddress, |
| 622 | uint64_t EndAddress); |
| 623 | |
| 624 | void computeInlinedContextSizeForFunc(const BinaryFunction *Func); |
| 625 | |
| 626 | void loadSymbolsFromPseudoProbe(); |
| 627 | |
| 628 | StringRef findPseudoProbeName(const BinaryFunction *Func); |
| 629 | |
| 630 | const MCDecodedPseudoProbe *getCallProbeForAddr(uint64_t Address) const { |
| 631 | return ProbeDecoder.getCallProbeForAddr(Address); |
| 632 | } |
| 633 | |
| 634 | void getInlineContextForProbe(const MCDecodedPseudoProbe *Probe, |
| 635 | SampleContextFrameVector &InlineContextStack, |
| 636 | bool IncludeLeaf = false) const { |
| 637 | SmallVector<MCPseudoProbeFrameLocation, 16> ProbeInlineContext; |
| 638 | ProbeDecoder.getInlineContextForProbe(Probe, InlineContextStack&: ProbeInlineContext, |
| 639 | IncludeLeaf); |
| 640 | for (uint32_t I = 0; I < ProbeInlineContext.size(); I++) { |
| 641 | auto &Callsite = ProbeInlineContext[I]; |
| 642 | // Clear the current context for an unknown probe. |
| 643 | if (Callsite.second == 0 && I != ProbeInlineContext.size() - 1) { |
| 644 | InlineContextStack.clear(); |
| 645 | continue; |
| 646 | } |
| 647 | InlineContextStack.emplace_back(Args: FunctionId(Callsite.first), |
| 648 | Args: LineLocation(Callsite.second, 0)); |
| 649 | } |
| 650 | } |
| 651 | const AddressProbesMap &getAddress2ProbesMap() const { |
| 652 | return ProbeDecoder.getAddress2ProbesMap(); |
| 653 | } |
| 654 | const MCPseudoProbeFuncDesc *getFuncDescForGUID(uint64_t GUID) { |
| 655 | return ProbeDecoder.getFuncDescForGUID(GUID); |
| 656 | } |
| 657 | |
| 658 | const MCPseudoProbeFuncDesc * |
| 659 | getInlinerDescForProbe(const MCDecodedPseudoProbe *Probe) { |
| 660 | return ProbeDecoder.getInlinerDescForProbe(Probe); |
| 661 | } |
| 662 | |
| 663 | bool isNonOverlappingAddressInterval(std::pair<uint64_t, uint64_t> LHS, |
| 664 | std::pair<uint64_t, uint64_t> RHS) { |
| 665 | if (LHS.second <= RHS.first || RHS.second <= LHS.first) |
| 666 | return true; |
| 667 | return false; |
| 668 | } |
| 669 | |
| 670 | Error addMMapNonTextEvent(MMapEvent Event) { |
| 671 | // Given the mmap events of the profiled binary, the virtual address |
| 672 | // intervals of mmaps most often doesn't overlap with each other. The |
| 673 | // implementation validates so, and runtime data address is mapped to |
| 674 | // a mmap event using look-up. With this implementation, data addresses |
| 675 | // from dynamic shared libraries (not the profiled binary) are not mapped or |
| 676 | // symbolized. To map runtime address to binary address in case of |
| 677 | // overlapping mmap events, the implementation could store all the mmap |
| 678 | // events in a vector and in the order they are added and reverse iterate |
| 679 | // the vector to find the mmap events. We opt'ed for the non-overlapping |
| 680 | // implementation for simplicity. |
| 681 | for (const auto &ExistingMMap : NonTextMMapEvents) { |
| 682 | if (isNonOverlappingAddressInterval( |
| 683 | LHS: {ExistingMMap.second.Address, |
| 684 | ExistingMMap.second.Address + ExistingMMap.second.Size}, |
| 685 | RHS: {Event.Address, Event.Address + Event.Size})) { |
| 686 | continue; |
| 687 | } |
| 688 | return createStringError( |
| 689 | EC: inconvertibleErrorCode(), |
| 690 | Fmt: "Non-text mmap event overlaps with existing event at address: %lx" , |
| 691 | Vals: Event.Address); |
| 692 | } |
| 693 | NonTextMMapEvents[Event.Address] = Event; |
| 694 | return Error::success(); |
| 695 | } |
| 696 | |
| 697 | // Given a non-text runtime address, canonicalize it to the virtual address in |
| 698 | // the binary. |
| 699 | // TODO: Consider unifying the canonicalization of text and non-text addresses |
| 700 | // in the ProfiledBinary class. |
| 701 | uint64_t CanonicalizeNonTextAddress(uint64_t Address); |
| 702 | |
| 703 | bool getTrackFuncContextSize() { return TrackFuncContextSize; } |
| 704 | |
| 705 | bool getIsLoadedByMMap() { return IsLoadedByMMap; } |
| 706 | |
| 707 | void setIsLoadedByMMap(bool Value) { IsLoadedByMMap = Value; } |
| 708 | |
| 709 | bool getMissingMMapWarned() { return MissingMMapWarned; } |
| 710 | |
| 711 | void setMissingMMapWarned(bool Value) { MissingMMapWarned = Value; } |
| 712 | }; |
| 713 | |
| 714 | } // end namespace sampleprof |
| 715 | } // end namespace llvm |
| 716 | |
| 717 | #endif |
| 718 | |