| 1 | //===- InstrProf.cpp - Instrumented profiling format support --------------===// |
| 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 | // This file contains support for clang's instrumentation based PGO and |
| 10 | // coverage. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/ProfileData/InstrProf.h" |
| 15 | #include "llvm/ADT/ArrayRef.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/ADT/StringExtras.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/Config/config.h" |
| 20 | #include "llvm/IR/Constant.h" |
| 21 | #include "llvm/IR/Constants.h" |
| 22 | #include "llvm/IR/Function.h" |
| 23 | #include "llvm/IR/GlobalValue.h" |
| 24 | #include "llvm/IR/GlobalVariable.h" |
| 25 | #include "llvm/IR/Instruction.h" |
| 26 | #include "llvm/IR/LLVMContext.h" |
| 27 | #include "llvm/IR/MDBuilder.h" |
| 28 | #include "llvm/IR/Metadata.h" |
| 29 | #include "llvm/IR/Module.h" |
| 30 | #include "llvm/IR/ProfDataUtils.h" |
| 31 | #include "llvm/IR/Type.h" |
| 32 | #include "llvm/ProfileData/InstrProfReader.h" |
| 33 | #include "llvm/Support/Casting.h" |
| 34 | #include "llvm/Support/CommandLine.h" |
| 35 | #include "llvm/Support/Compiler.h" |
| 36 | #include "llvm/Support/Compression.h" |
| 37 | #include "llvm/Support/Debug.h" |
| 38 | #include "llvm/Support/Endian.h" |
| 39 | #include "llvm/Support/Error.h" |
| 40 | #include "llvm/Support/ErrorHandling.h" |
| 41 | #include "llvm/Support/LEB128.h" |
| 42 | #include "llvm/Support/MathExtras.h" |
| 43 | #include "llvm/Support/Path.h" |
| 44 | #include "llvm/Support/SwapByteOrder.h" |
| 45 | #include "llvm/Support/VirtualFileSystem.h" |
| 46 | #include "llvm/Support/raw_ostream.h" |
| 47 | #include "llvm/TargetParser/Triple.h" |
| 48 | #include <algorithm> |
| 49 | #include <cassert> |
| 50 | #include <cstddef> |
| 51 | #include <cstdint> |
| 52 | #include <cstring> |
| 53 | #include <memory> |
| 54 | #include <string> |
| 55 | #include <system_error> |
| 56 | #include <type_traits> |
| 57 | #include <utility> |
| 58 | #include <vector> |
| 59 | |
| 60 | using namespace llvm; |
| 61 | |
| 62 | #define DEBUG_TYPE "instrprof" |
| 63 | |
| 64 | static cl::opt<bool> StaticFuncFullModulePrefix( |
| 65 | "static-func-full-module-prefix" , cl::init(Val: true), cl::Hidden, |
| 66 | cl::desc("Use full module build paths in the profile counter names for " |
| 67 | "static functions." )); |
| 68 | |
| 69 | // This option is tailored to users that have different top-level directory in |
| 70 | // profile-gen and profile-use compilation. Users need to specific the number |
| 71 | // of levels to strip. A value larger than the number of directories in the |
| 72 | // source file will strip all the directory names and only leave the basename. |
| 73 | // |
| 74 | // Note current ThinLTO module importing for the indirect-calls assumes |
| 75 | // the source directory name not being stripped. A non-zero option value here |
| 76 | // can potentially prevent some inter-module indirect-call-promotions. |
| 77 | static cl::opt<unsigned> StaticFuncStripDirNamePrefix( |
| 78 | "static-func-strip-dirname-prefix" , cl::init(Val: 0), cl::Hidden, |
| 79 | cl::desc("Strip specified level of directory name from source path in " |
| 80 | "the profile counter name for static functions." )); |
| 81 | |
| 82 | static std::string getInstrProfErrString(instrprof_error Err, |
| 83 | const std::string &ErrMsg = "" ) { |
| 84 | std::string Msg; |
| 85 | raw_string_ostream OS(Msg); |
| 86 | |
| 87 | switch (Err) { |
| 88 | case instrprof_error::success: |
| 89 | OS << "success" ; |
| 90 | break; |
| 91 | case instrprof_error::eof: |
| 92 | OS << "end of File" ; |
| 93 | break; |
| 94 | case instrprof_error::unrecognized_format: |
| 95 | OS << "unrecognized instrumentation profile encoding format" ; |
| 96 | break; |
| 97 | case instrprof_error::bad_magic: |
| 98 | OS << "invalid instrumentation profile data (bad magic)" ; |
| 99 | break; |
| 100 | case instrprof_error::bad_header: |
| 101 | OS << "invalid instrumentation profile data (file header is corrupt)" ; |
| 102 | break; |
| 103 | case instrprof_error::unsupported_version: |
| 104 | OS << "unsupported instrumentation profile format version" ; |
| 105 | break; |
| 106 | case instrprof_error::unsupported_hash_type: |
| 107 | OS << "unsupported instrumentation profile hash type" ; |
| 108 | break; |
| 109 | case instrprof_error::too_large: |
| 110 | OS << "too much profile data" ; |
| 111 | break; |
| 112 | case instrprof_error::truncated: |
| 113 | OS << "truncated profile data" ; |
| 114 | break; |
| 115 | case instrprof_error::malformed: |
| 116 | OS << "malformed instrumentation profile data" ; |
| 117 | break; |
| 118 | case instrprof_error::missing_correlation_info: |
| 119 | OS << "debug info/binary for correlation is required" ; |
| 120 | break; |
| 121 | case instrprof_error::unexpected_correlation_info: |
| 122 | OS << "debug info/binary for correlation is not necessary" ; |
| 123 | break; |
| 124 | case instrprof_error::unable_to_correlate_profile: |
| 125 | OS << "unable to correlate profile" ; |
| 126 | break; |
| 127 | case instrprof_error::invalid_prof: |
| 128 | OS << "invalid profile created. Please file a bug " |
| 129 | "at: " BUG_REPORT_URL |
| 130 | " and include the profraw files that caused this error." ; |
| 131 | break; |
| 132 | case instrprof_error::unknown_function: |
| 133 | OS << "no profile data available for function" ; |
| 134 | break; |
| 135 | case instrprof_error::hash_mismatch: |
| 136 | OS << "function control flow change detected (hash mismatch)" ; |
| 137 | break; |
| 138 | case instrprof_error::count_mismatch: |
| 139 | OS << "function basic block count change detected (counter mismatch)" ; |
| 140 | break; |
| 141 | case instrprof_error::bitmap_mismatch: |
| 142 | OS << "function bitmap size change detected (bitmap size mismatch)" ; |
| 143 | break; |
| 144 | case instrprof_error::counter_overflow: |
| 145 | OS << "counter overflow" ; |
| 146 | break; |
| 147 | case instrprof_error::value_site_count_mismatch: |
| 148 | OS << "function value site count change detected (counter mismatch)" ; |
| 149 | break; |
| 150 | case instrprof_error::compress_failed: |
| 151 | OS << "failed to compress data (zlib)" ; |
| 152 | break; |
| 153 | case instrprof_error::uncompress_failed: |
| 154 | OS << "failed to uncompress data (zlib)" ; |
| 155 | break; |
| 156 | case instrprof_error::empty_raw_profile: |
| 157 | OS << "empty raw profile file" ; |
| 158 | break; |
| 159 | case instrprof_error::zlib_unavailable: |
| 160 | OS << "profile uses zlib compression but the profile reader was built " |
| 161 | "without zlib support" ; |
| 162 | break; |
| 163 | case instrprof_error::raw_profile_version_mismatch: |
| 164 | OS << "raw profile version mismatch" ; |
| 165 | break; |
| 166 | case instrprof_error::counter_value_too_large: |
| 167 | OS << "excessively large counter value suggests corrupted profile data" ; |
| 168 | break; |
| 169 | } |
| 170 | |
| 171 | // If optional error message is not empty, append it to the message. |
| 172 | if (!ErrMsg.empty()) |
| 173 | OS << ": " << ErrMsg; |
| 174 | |
| 175 | return OS.str(); |
| 176 | } |
| 177 | |
| 178 | namespace { |
| 179 | |
| 180 | // FIXME: This class is only here to support the transition to llvm::Error. It |
| 181 | // will be removed once this transition is complete. Clients should prefer to |
| 182 | // deal with the Error value directly, rather than converting to error_code. |
| 183 | class InstrProfErrorCategoryType : public std::error_category { |
| 184 | const char *name() const noexcept override { return "llvm.instrprof" ; } |
| 185 | |
| 186 | std::string message(int IE) const override { |
| 187 | return getInstrProfErrString(Err: static_cast<instrprof_error>(IE)); |
| 188 | } |
| 189 | }; |
| 190 | |
| 191 | } // end anonymous namespace |
| 192 | |
| 193 | const std::error_category &llvm::instrprof_category() { |
| 194 | static InstrProfErrorCategoryType ErrorCategory; |
| 195 | return ErrorCategory; |
| 196 | } |
| 197 | |
| 198 | namespace { |
| 199 | |
| 200 | const char *InstrProfSectNameCommon[] = { |
| 201 | #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ |
| 202 | SectNameCommon, |
| 203 | #include "llvm/ProfileData/InstrProfData.inc" |
| 204 | }; |
| 205 | |
| 206 | const char *InstrProfSectNameCoff[] = { |
| 207 | #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ |
| 208 | SectNameCoff, |
| 209 | #include "llvm/ProfileData/InstrProfData.inc" |
| 210 | }; |
| 211 | |
| 212 | const char *InstrProfSectNamePrefix[] = { |
| 213 | #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ |
| 214 | Prefix, |
| 215 | #include "llvm/ProfileData/InstrProfData.inc" |
| 216 | }; |
| 217 | |
| 218 | } // namespace |
| 219 | |
| 220 | namespace llvm { |
| 221 | |
| 222 | cl::opt<bool> DoInstrProfNameCompression( |
| 223 | "enable-name-compression" , |
| 224 | cl::desc("Enable name/filename string compression" ), cl::init(Val: true)); |
| 225 | |
| 226 | cl::opt<bool> EnableVTableValueProfiling( |
| 227 | "enable-vtable-value-profiling" , cl::init(Val: false), |
| 228 | cl::desc("If true, the virtual table address will be instrumented to know " |
| 229 | "the types of a C++ pointer. The information is used in indirect " |
| 230 | "call promotion to do selective vtable-based comparison." )); |
| 231 | |
| 232 | cl::opt<bool> EnableVTableProfileUse( |
| 233 | "enable-vtable-profile-use" , cl::init(Val: false), |
| 234 | cl::desc("If ThinLTO and WPD is enabled and this option is true, vtable " |
| 235 | "profiles will be used by ICP pass for more efficient indirect " |
| 236 | "call sequence. If false, type profiles won't be used." )); |
| 237 | |
| 238 | std::string getInstrProfSectionName(InstrProfSectKind IPSK, |
| 239 | Triple::ObjectFormatType OF, |
| 240 | bool AddSegmentInfo) { |
| 241 | std::string SectName; |
| 242 | |
| 243 | if (OF == Triple::MachO && AddSegmentInfo) |
| 244 | SectName = InstrProfSectNamePrefix[IPSK]; |
| 245 | |
| 246 | if (OF == Triple::COFF) |
| 247 | SectName += InstrProfSectNameCoff[IPSK]; |
| 248 | else |
| 249 | SectName += InstrProfSectNameCommon[IPSK]; |
| 250 | |
| 251 | if (OF == Triple::MachO && IPSK == IPSK_data && AddSegmentInfo) |
| 252 | SectName += ",regular,live_support" ; |
| 253 | |
| 254 | return SectName; |
| 255 | } |
| 256 | |
| 257 | std::string InstrProfError::message() const { |
| 258 | return getInstrProfErrString(Err, ErrMsg: Msg); |
| 259 | } |
| 260 | |
| 261 | char InstrProfError::ID = 0; |
| 262 | |
| 263 | ProfOStream::ProfOStream(raw_fd_ostream &FD) |
| 264 | : IsFDOStream(true), OS(FD), LE(FD, llvm::endianness::little) {} |
| 265 | |
| 266 | ProfOStream::ProfOStream(raw_string_ostream &STR) |
| 267 | : IsFDOStream(false), OS(STR), LE(STR, llvm::endianness::little) {} |
| 268 | |
| 269 | uint64_t ProfOStream::tell() const { return OS.tell(); } |
| 270 | void ProfOStream::write(uint64_t V) { LE.write<uint64_t>(Val: V); } |
| 271 | void ProfOStream::write32(uint32_t V) { LE.write<uint32_t>(Val: V); } |
| 272 | void ProfOStream::writeByte(uint8_t V) { LE.write<uint8_t>(Val: V); } |
| 273 | |
| 274 | void ProfOStream::patch(ArrayRef<PatchItem> P) { |
| 275 | using namespace support; |
| 276 | |
| 277 | if (IsFDOStream) { |
| 278 | raw_fd_ostream &FDOStream = static_cast<raw_fd_ostream &>(OS); |
| 279 | const uint64_t LastPos = FDOStream.tell(); |
| 280 | for (const auto &K : P) { |
| 281 | FDOStream.seek(off: K.Pos); |
| 282 | for (uint64_t Elem : K.D) |
| 283 | write(V: Elem); |
| 284 | } |
| 285 | // Reset the stream to the last position after patching so that users |
| 286 | // don't accidentally overwrite data. This makes it consistent with |
| 287 | // the string stream below which replaces the data directly. |
| 288 | FDOStream.seek(off: LastPos); |
| 289 | } else { |
| 290 | raw_string_ostream &SOStream = static_cast<raw_string_ostream &>(OS); |
| 291 | std::string &Data = SOStream.str(); // with flush |
| 292 | for (const auto &K : P) { |
| 293 | for (int I = 0, E = K.D.size(); I != E; I++) { |
| 294 | uint64_t Bytes = |
| 295 | endian::byte_swap<uint64_t, llvm::endianness::little>(value: K.D[I]); |
| 296 | Data.replace(pos: K.Pos + I * sizeof(uint64_t), n1: sizeof(uint64_t), |
| 297 | s: (const char *)&Bytes, n2: sizeof(uint64_t)); |
| 298 | } |
| 299 | } |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | std::string getPGOFuncName(StringRef Name, GlobalValue::LinkageTypes Linkage, |
| 304 | StringRef FileName, |
| 305 | uint64_t Version LLVM_ATTRIBUTE_UNUSED) { |
| 306 | // Value names may be prefixed with a binary '1' to indicate |
| 307 | // that the backend should not modify the symbols due to any platform |
| 308 | // naming convention. Do not include that '1' in the PGO profile name. |
| 309 | if (Name[0] == '\1') |
| 310 | Name = Name.substr(Start: 1); |
| 311 | |
| 312 | std::string NewName = std::string(Name); |
| 313 | if (llvm::GlobalValue::isLocalLinkage(Linkage)) { |
| 314 | // For local symbols, prepend the main file name to distinguish them. |
| 315 | // Do not include the full path in the file name since there's no guarantee |
| 316 | // that it will stay the same, e.g., if the files are checked out from |
| 317 | // version control in different locations. |
| 318 | if (FileName.empty()) |
| 319 | NewName = NewName.insert(pos: 0, s: "<unknown>:" ); |
| 320 | else |
| 321 | NewName = NewName.insert(pos1: 0, str: FileName.str() + ":" ); |
| 322 | } |
| 323 | return NewName; |
| 324 | } |
| 325 | |
| 326 | // Strip NumPrefix level of directory name from PathNameStr. If the number of |
| 327 | // directory separators is less than NumPrefix, strip all the directories and |
| 328 | // leave base file name only. |
| 329 | static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) { |
| 330 | uint32_t Count = NumPrefix; |
| 331 | uint32_t Pos = 0, LastPos = 0; |
| 332 | for (const auto &CI : PathNameStr) { |
| 333 | ++Pos; |
| 334 | if (llvm::sys::path::is_separator(value: CI)) { |
| 335 | LastPos = Pos; |
| 336 | --Count; |
| 337 | } |
| 338 | if (Count == 0) |
| 339 | break; |
| 340 | } |
| 341 | return PathNameStr.substr(Start: LastPos); |
| 342 | } |
| 343 | |
| 344 | static StringRef getStrippedSourceFileName(const GlobalObject &GO) { |
| 345 | StringRef FileName(GO.getParent()->getSourceFileName()); |
| 346 | uint32_t StripLevel = StaticFuncFullModulePrefix ? 0 : (uint32_t)-1; |
| 347 | if (StripLevel < StaticFuncStripDirNamePrefix) |
| 348 | StripLevel = StaticFuncStripDirNamePrefix; |
| 349 | if (StripLevel) |
| 350 | FileName = stripDirPrefix(PathNameStr: FileName, NumPrefix: StripLevel); |
| 351 | return FileName; |
| 352 | } |
| 353 | |
| 354 | // The PGO name has the format [<filepath>;]<mangled-name> where <filepath>; is |
| 355 | // provided if linkage is local and is used to discriminate possibly identical |
| 356 | // mangled names. ";" is used because it is unlikely to be found in either |
| 357 | // <filepath> or <mangled-name>. |
| 358 | // |
| 359 | // Older compilers used getPGOFuncName() which has the format |
| 360 | // [<filepath>:]<mangled-name>. This caused trouble for Objective-C functions |
| 361 | // which commonly have :'s in their names. We still need to compute this name to |
| 362 | // lookup functions from profiles built by older compilers. |
| 363 | static std::string |
| 364 | getIRPGONameForGlobalObject(const GlobalObject &GO, |
| 365 | GlobalValue::LinkageTypes Linkage, |
| 366 | StringRef FileName) { |
| 367 | return GlobalValue::getGlobalIdentifier(Name: GO.getName(), Linkage, FileName); |
| 368 | } |
| 369 | |
| 370 | static std::optional<std::string> lookupPGONameFromMetadata(MDNode *MD) { |
| 371 | if (MD != nullptr) { |
| 372 | StringRef S = cast<MDString>(Val: MD->getOperand(I: 0))->getString(); |
| 373 | return S.str(); |
| 374 | } |
| 375 | return {}; |
| 376 | } |
| 377 | |
| 378 | // Returns the PGO object name. This function has some special handling |
| 379 | // when called in LTO optimization. The following only applies when calling in |
| 380 | // LTO passes (when \c InLTO is true): LTO's internalization privatizes many |
| 381 | // global linkage symbols. This happens after value profile annotation, but |
| 382 | // those internal linkage functions should not have a source prefix. |
| 383 | // Additionally, for ThinLTO mode, exported internal functions are promoted |
| 384 | // and renamed. We need to ensure that the original internal PGO name is |
| 385 | // used when computing the GUID that is compared against the profiled GUIDs. |
| 386 | // To differentiate compiler generated internal symbols from original ones, |
| 387 | // PGOFuncName meta data are created and attached to the original internal |
| 388 | // symbols in the value profile annotation step |
| 389 | // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta |
| 390 | // data, its original linkage must be non-internal. |
| 391 | static std::string getIRPGOObjectName(const GlobalObject &GO, bool InLTO, |
| 392 | MDNode *PGONameMetadata) { |
| 393 | if (!InLTO) { |
| 394 | auto FileName = getStrippedSourceFileName(GO); |
| 395 | return getIRPGONameForGlobalObject(GO, Linkage: GO.getLinkage(), FileName); |
| 396 | } |
| 397 | |
| 398 | // In LTO mode (when InLTO is true), first check if there is a meta data. |
| 399 | if (auto IRPGOFuncName = lookupPGONameFromMetadata(MD: PGONameMetadata)) |
| 400 | return *IRPGOFuncName; |
| 401 | |
| 402 | // If there is no meta data, the function must be a global before the value |
| 403 | // profile annotation pass. Its current linkage may be internal if it is |
| 404 | // internalized in LTO mode. |
| 405 | return getIRPGONameForGlobalObject(GO, Linkage: GlobalValue::ExternalLinkage, FileName: "" ); |
| 406 | } |
| 407 | |
| 408 | // Returns the IRPGO function name and does special handling when called |
| 409 | // in LTO optimization. See the comments of `getIRPGOObjectName` for details. |
| 410 | std::string getIRPGOFuncName(const Function &F, bool InLTO) { |
| 411 | return getIRPGOObjectName(GO: F, InLTO, PGONameMetadata: getPGOFuncNameMetadata(F)); |
| 412 | } |
| 413 | |
| 414 | // Please use getIRPGOFuncName for LLVM IR instrumentation. This function is |
| 415 | // for front-end (Clang, etc) instrumentation. |
| 416 | // The implementation is kept for profile matching from older profiles. |
| 417 | // This is similar to `getIRPGOFuncName` except that this function calls |
| 418 | // 'getPGOFuncName' to get a name and `getIRPGOFuncName` calls |
| 419 | // 'getIRPGONameForGlobalObject'. See the difference between two callees in the |
| 420 | // comments of `getIRPGONameForGlobalObject`. |
| 421 | std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) { |
| 422 | if (!InLTO) { |
| 423 | auto FileName = getStrippedSourceFileName(GO: F); |
| 424 | return getPGOFuncName(Name: F.getName(), Linkage: F.getLinkage(), FileName, Version); |
| 425 | } |
| 426 | |
| 427 | // In LTO mode (when InLTO is true), first check if there is a meta data. |
| 428 | if (auto PGOFuncName = lookupPGONameFromMetadata(MD: getPGOFuncNameMetadata(F))) |
| 429 | return *PGOFuncName; |
| 430 | |
| 431 | // If there is no meta data, the function must be a global before the value |
| 432 | // profile annotation pass. Its current linkage may be internal if it is |
| 433 | // internalized in LTO mode. |
| 434 | return getPGOFuncName(Name: F.getName(), Linkage: GlobalValue::ExternalLinkage, FileName: "" ); |
| 435 | } |
| 436 | |
| 437 | std::string getPGOName(const GlobalVariable &V, bool InLTO) { |
| 438 | // PGONameMetadata should be set by compiler at profile use time |
| 439 | // and read by symtab creation to look up symbols corresponding to |
| 440 | // a MD5 hash. |
| 441 | return getIRPGOObjectName(GO: V, InLTO, PGONameMetadata: V.getMetadata(Kind: getPGONameMetadataName())); |
| 442 | } |
| 443 | |
| 444 | // See getIRPGOObjectName() for a discription of the format. |
| 445 | std::pair<StringRef, StringRef> getParsedIRPGOName(StringRef IRPGOName) { |
| 446 | auto [FileName, MangledName] = IRPGOName.split(Separator: GlobalIdentifierDelimiter); |
| 447 | if (MangledName.empty()) |
| 448 | return std::make_pair(x: StringRef(), y&: IRPGOName); |
| 449 | return std::make_pair(x&: FileName, y&: MangledName); |
| 450 | } |
| 451 | |
| 452 | StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) { |
| 453 | if (FileName.empty()) |
| 454 | return PGOFuncName; |
| 455 | // Drop the file name including ':' or ';'. See getIRPGONameForGlobalObject as |
| 456 | // well. |
| 457 | if (PGOFuncName.starts_with(Prefix: FileName)) |
| 458 | PGOFuncName = PGOFuncName.drop_front(N: FileName.size() + 1); |
| 459 | return PGOFuncName; |
| 460 | } |
| 461 | |
| 462 | // \p FuncName is the string used as profile lookup key for the function. A |
| 463 | // symbol is created to hold the name. Return the legalized symbol name. |
| 464 | std::string getPGOFuncNameVarName(StringRef FuncName, |
| 465 | GlobalValue::LinkageTypes Linkage) { |
| 466 | std::string VarName = std::string(getInstrProfNameVarPrefix()); |
| 467 | VarName += FuncName; |
| 468 | |
| 469 | if (!GlobalValue::isLocalLinkage(Linkage)) |
| 470 | return VarName; |
| 471 | |
| 472 | // Now fix up illegal chars in local VarName that may upset the assembler. |
| 473 | const char InvalidChars[] = "-:;<>/\"'" ; |
| 474 | size_t FoundPos = VarName.find_first_of(s: InvalidChars); |
| 475 | while (FoundPos != std::string::npos) { |
| 476 | VarName[FoundPos] = '_'; |
| 477 | FoundPos = VarName.find_first_of(s: InvalidChars, pos: FoundPos + 1); |
| 478 | } |
| 479 | return VarName; |
| 480 | } |
| 481 | |
| 482 | bool isGPUProfTarget(const Module &M) { |
| 483 | const Triple &T = M.getTargetTriple(); |
| 484 | return T.isGPU(); |
| 485 | } |
| 486 | |
| 487 | void setPGOFuncVisibility(Module &M, GlobalVariable *FuncNameVar) { |
| 488 | // If the target is a GPU, make the symbol protected so it can |
| 489 | // be read from the host device |
| 490 | if (isGPUProfTarget(M)) |
| 491 | FuncNameVar->setVisibility(GlobalValue::ProtectedVisibility); |
| 492 | // Hide the symbol so that we correctly get a copy for each executable. |
| 493 | else if (!GlobalValue::isLocalLinkage(Linkage: FuncNameVar->getLinkage())) |
| 494 | FuncNameVar->setVisibility(GlobalValue::HiddenVisibility); |
| 495 | } |
| 496 | |
| 497 | GlobalVariable *createPGOFuncNameVar(Module &M, |
| 498 | GlobalValue::LinkageTypes Linkage, |
| 499 | StringRef PGOFuncName) { |
| 500 | // Ensure profiling variables on GPU are visible to be read from host |
| 501 | if (isGPUProfTarget(M)) |
| 502 | Linkage = GlobalValue::ExternalLinkage; |
| 503 | // We generally want to match the function's linkage, but available_externally |
| 504 | // and extern_weak both have the wrong semantics, and anything that doesn't |
| 505 | // need to link across compilation units doesn't need to be visible at all. |
| 506 | else if (Linkage == GlobalValue::ExternalWeakLinkage) |
| 507 | Linkage = GlobalValue::LinkOnceAnyLinkage; |
| 508 | else if (Linkage == GlobalValue::AvailableExternallyLinkage) |
| 509 | Linkage = GlobalValue::LinkOnceODRLinkage; |
| 510 | else if (Linkage == GlobalValue::InternalLinkage || |
| 511 | Linkage == GlobalValue::ExternalLinkage) |
| 512 | Linkage = GlobalValue::PrivateLinkage; |
| 513 | |
| 514 | auto *Value = |
| 515 | ConstantDataArray::getString(Context&: M.getContext(), Initializer: PGOFuncName, AddNull: false); |
| 516 | auto *FuncNameVar = |
| 517 | new GlobalVariable(M, Value->getType(), true, Linkage, Value, |
| 518 | getPGOFuncNameVarName(FuncName: PGOFuncName, Linkage)); |
| 519 | |
| 520 | setPGOFuncVisibility(M, FuncNameVar); |
| 521 | return FuncNameVar; |
| 522 | } |
| 523 | |
| 524 | GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) { |
| 525 | return createPGOFuncNameVar(M&: *F.getParent(), Linkage: F.getLinkage(), PGOFuncName); |
| 526 | } |
| 527 | |
| 528 | Error InstrProfSymtab::create(Module &M, bool InLTO, bool AddCanonical) { |
| 529 | for (Function &F : M) { |
| 530 | // Function may not have a name: like using asm("") to overwrite the name. |
| 531 | // Ignore in this case. |
| 532 | if (!F.hasName()) |
| 533 | continue; |
| 534 | if (Error E = addFuncWithName(F, PGOFuncName: getIRPGOFuncName(F, InLTO), AddCanonical)) |
| 535 | return E; |
| 536 | // Also use getPGOFuncName() so that we can find records from older profiles |
| 537 | if (Error E = addFuncWithName(F, PGOFuncName: getPGOFuncName(F, InLTO), AddCanonical)) |
| 538 | return E; |
| 539 | } |
| 540 | |
| 541 | for (GlobalVariable &G : M.globals()) { |
| 542 | if (!G.hasName() || !G.hasMetadata(KindID: LLVMContext::MD_type)) |
| 543 | continue; |
| 544 | if (Error E = addVTableWithName(V&: G, PGOVTableName: getPGOName(V: G, InLTO))) |
| 545 | return E; |
| 546 | } |
| 547 | |
| 548 | Sorted = false; |
| 549 | finalizeSymtab(); |
| 550 | return Error::success(); |
| 551 | } |
| 552 | |
| 553 | Error InstrProfSymtab::addVTableWithName(GlobalVariable &VTable, |
| 554 | StringRef VTablePGOName) { |
| 555 | auto NameToGUIDMap = [&](StringRef Name) -> Error { |
| 556 | if (Error E = addSymbolName(SymbolName: Name)) |
| 557 | return E; |
| 558 | |
| 559 | bool Inserted = true; |
| 560 | std::tie(args: std::ignore, args&: Inserted) = MD5VTableMap.try_emplace( |
| 561 | Key: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: Name), Args: &VTable); |
| 562 | if (!Inserted) |
| 563 | LLVM_DEBUG(dbgs() << "GUID conflict within one module" ); |
| 564 | return Error::success(); |
| 565 | }; |
| 566 | if (Error E = NameToGUIDMap(VTablePGOName)) |
| 567 | return E; |
| 568 | |
| 569 | StringRef CanonicalName = getCanonicalName(PGOName: VTablePGOName); |
| 570 | if (CanonicalName != VTablePGOName) |
| 571 | return NameToGUIDMap(CanonicalName); |
| 572 | |
| 573 | return Error::success(); |
| 574 | } |
| 575 | |
| 576 | Error readAndDecodeStrings(StringRef NameStrings, |
| 577 | std::function<Error(StringRef)> NameCallback) { |
| 578 | const uint8_t *P = NameStrings.bytes_begin(); |
| 579 | const uint8_t *EndP = NameStrings.bytes_end(); |
| 580 | while (P < EndP) { |
| 581 | uint32_t N; |
| 582 | uint64_t UncompressedSize = decodeULEB128(p: P, n: &N); |
| 583 | P += N; |
| 584 | uint64_t CompressedSize = decodeULEB128(p: P, n: &N); |
| 585 | P += N; |
| 586 | const bool IsCompressed = (CompressedSize != 0); |
| 587 | SmallVector<uint8_t, 128> UncompressedNameStrings; |
| 588 | StringRef NameStrings; |
| 589 | if (IsCompressed) { |
| 590 | if (!llvm::compression::zlib::isAvailable()) |
| 591 | return make_error<InstrProfError>(Args: instrprof_error::zlib_unavailable); |
| 592 | |
| 593 | if (Error E = compression::zlib::decompress(Input: ArrayRef(P, CompressedSize), |
| 594 | Output&: UncompressedNameStrings, |
| 595 | UncompressedSize)) { |
| 596 | consumeError(Err: std::move(E)); |
| 597 | return make_error<InstrProfError>(Args: instrprof_error::uncompress_failed); |
| 598 | } |
| 599 | P += CompressedSize; |
| 600 | NameStrings = toStringRef(Input: UncompressedNameStrings); |
| 601 | } else { |
| 602 | NameStrings = |
| 603 | StringRef(reinterpret_cast<const char *>(P), UncompressedSize); |
| 604 | P += UncompressedSize; |
| 605 | } |
| 606 | // Now parse the name strings. |
| 607 | SmallVector<StringRef, 0> Names; |
| 608 | NameStrings.split(A&: Names, Separator: getInstrProfNameSeparator()); |
| 609 | for (StringRef &Name : Names) |
| 610 | if (Error E = NameCallback(Name)) |
| 611 | return E; |
| 612 | |
| 613 | while (P < EndP && *P == 0) |
| 614 | P++; |
| 615 | } |
| 616 | return Error::success(); |
| 617 | } |
| 618 | |
| 619 | Error InstrProfSymtab::create(StringRef NameStrings) { |
| 620 | return readAndDecodeStrings(NameStrings, |
| 621 | NameCallback: [&](StringRef S) { return addFuncName(FuncName: S); }); |
| 622 | } |
| 623 | |
| 624 | Error InstrProfSymtab::create(StringRef FuncNameStrings, |
| 625 | StringRef VTableNameStrings) { |
| 626 | if (Error E = readAndDecodeStrings( |
| 627 | NameStrings: FuncNameStrings, NameCallback: [&](StringRef S) { return addFuncName(FuncName: S); })) |
| 628 | return E; |
| 629 | |
| 630 | return readAndDecodeStrings(NameStrings: VTableNameStrings, |
| 631 | NameCallback: [&](StringRef S) { return addVTableName(VTableName: S); }); |
| 632 | } |
| 633 | |
| 634 | Error InstrProfSymtab::initVTableNamesFromCompressedStrings( |
| 635 | StringRef CompressedVTableStrings) { |
| 636 | return readAndDecodeStrings(NameStrings: CompressedVTableStrings, |
| 637 | NameCallback: [&](StringRef S) { return addVTableName(VTableName: S); }); |
| 638 | } |
| 639 | |
| 640 | StringRef InstrProfSymtab::getCanonicalName(StringRef PGOName) { |
| 641 | // In ThinLTO, local function may have been promoted to global and have |
| 642 | // suffix ".llvm." added to the function name. We need to add the |
| 643 | // stripped function name to the symbol table so that we can find a match |
| 644 | // from profile. |
| 645 | // |
| 646 | // ".__uniq." suffix is used to differentiate internal linkage functions in |
| 647 | // different modules and should be kept. This is the only suffix with the |
| 648 | // pattern ".xxx" which is kept before matching, other suffixes similar as |
| 649 | // ".llvm." will be stripped. |
| 650 | const std::string UniqSuffix = ".__uniq." ; |
| 651 | size_t Pos = PGOName.find(Str: UniqSuffix); |
| 652 | if (Pos != StringRef::npos) |
| 653 | Pos += UniqSuffix.length(); |
| 654 | else |
| 655 | Pos = 0; |
| 656 | |
| 657 | // Search '.' after ".__uniq." if ".__uniq." exists, otherwise search '.' from |
| 658 | // the beginning. |
| 659 | Pos = PGOName.find(C: '.', From: Pos); |
| 660 | if (Pos != StringRef::npos && Pos != 0) |
| 661 | return PGOName.substr(Start: 0, N: Pos); |
| 662 | |
| 663 | return PGOName; |
| 664 | } |
| 665 | |
| 666 | Error InstrProfSymtab::addFuncWithName(Function &F, StringRef PGOFuncName, |
| 667 | bool AddCanonical) { |
| 668 | auto NameToGUIDMap = [&](StringRef Name) -> Error { |
| 669 | if (Error E = addFuncName(FuncName: Name)) |
| 670 | return E; |
| 671 | MD5FuncMap.emplace_back(args: Function::getGUIDAssumingExternalLinkage(GlobalName: Name), args: &F); |
| 672 | return Error::success(); |
| 673 | }; |
| 674 | if (Error E = NameToGUIDMap(PGOFuncName)) |
| 675 | return E; |
| 676 | |
| 677 | if (!AddCanonical) |
| 678 | return Error::success(); |
| 679 | |
| 680 | StringRef CanonicalFuncName = getCanonicalName(PGOName: PGOFuncName); |
| 681 | if (CanonicalFuncName != PGOFuncName) |
| 682 | return NameToGUIDMap(CanonicalFuncName); |
| 683 | |
| 684 | return Error::success(); |
| 685 | } |
| 686 | |
| 687 | uint64_t InstrProfSymtab::getVTableHashFromAddress(uint64_t Address) { |
| 688 | // Given a runtime address, look up the hash value in the interval map, and |
| 689 | // fallback to value 0 if a hash value is not found. |
| 690 | return VTableAddrMap.lookup(x: Address, NotFound: 0); |
| 691 | } |
| 692 | |
| 693 | uint64_t InstrProfSymtab::getFunctionHashFromAddress(uint64_t Address) { |
| 694 | finalizeSymtab(); |
| 695 | auto It = partition_point(Range&: AddrToMD5Map, P: [=](std::pair<uint64_t, uint64_t> A) { |
| 696 | return A.first < Address; |
| 697 | }); |
| 698 | // Raw function pointer collected by value profiler may be from |
| 699 | // external functions that are not instrumented. They won't have |
| 700 | // mapping data to be used by the deserializer. Force the value to |
| 701 | // be 0 in this case. |
| 702 | if (It != AddrToMD5Map.end() && It->first == Address) |
| 703 | return (uint64_t)It->second; |
| 704 | return 0; |
| 705 | } |
| 706 | |
| 707 | void InstrProfSymtab::dumpNames(raw_ostream &OS) const { |
| 708 | SmallVector<StringRef, 0> Sorted(NameTab.keys()); |
| 709 | llvm::sort(C&: Sorted); |
| 710 | for (StringRef S : Sorted) |
| 711 | OS << S << '\n'; |
| 712 | } |
| 713 | |
| 714 | Error collectGlobalObjectNameStrings(ArrayRef<std::string> NameStrs, |
| 715 | bool DoCompression, std::string &Result) { |
| 716 | assert(!NameStrs.empty() && "No name data to emit" ); |
| 717 | |
| 718 | uint8_t [20], *P = Header; |
| 719 | std::string UncompressedNameStrings = |
| 720 | join(Begin: NameStrs.begin(), End: NameStrs.end(), Separator: getInstrProfNameSeparator()); |
| 721 | |
| 722 | assert(StringRef(UncompressedNameStrings) |
| 723 | .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) && |
| 724 | "PGO name is invalid (contains separator token)" ); |
| 725 | |
| 726 | unsigned EncLen = encodeULEB128(Value: UncompressedNameStrings.length(), p: P); |
| 727 | P += EncLen; |
| 728 | |
| 729 | auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) { |
| 730 | EncLen = encodeULEB128(Value: CompressedLen, p: P); |
| 731 | P += EncLen; |
| 732 | char * = reinterpret_cast<char *>(&Header[0]); |
| 733 | unsigned = P - &Header[0]; |
| 734 | Result.append(s: HeaderStr, n: HeaderLen); |
| 735 | Result += InputStr; |
| 736 | return Error::success(); |
| 737 | }; |
| 738 | |
| 739 | if (!DoCompression) { |
| 740 | return WriteStringToResult(0, UncompressedNameStrings); |
| 741 | } |
| 742 | |
| 743 | SmallVector<uint8_t, 128> CompressedNameStrings; |
| 744 | compression::zlib::compress(Input: arrayRefFromStringRef(Input: UncompressedNameStrings), |
| 745 | CompressedBuffer&: CompressedNameStrings, |
| 746 | Level: compression::zlib::BestSizeCompression); |
| 747 | |
| 748 | return WriteStringToResult(CompressedNameStrings.size(), |
| 749 | toStringRef(Input: CompressedNameStrings)); |
| 750 | } |
| 751 | |
| 752 | StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) { |
| 753 | auto *Arr = cast<ConstantDataArray>(Val: NameVar->getInitializer()); |
| 754 | StringRef NameStr = |
| 755 | Arr->isCString() ? Arr->getAsCString() : Arr->getAsString(); |
| 756 | return NameStr; |
| 757 | } |
| 758 | |
| 759 | Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars, |
| 760 | std::string &Result, bool DoCompression) { |
| 761 | std::vector<std::string> NameStrs; |
| 762 | for (auto *NameVar : NameVars) { |
| 763 | NameStrs.push_back(x: std::string(getPGOFuncNameVarInitializer(NameVar))); |
| 764 | } |
| 765 | return collectGlobalObjectNameStrings( |
| 766 | NameStrs, DoCompression: compression::zlib::isAvailable() && DoCompression, Result); |
| 767 | } |
| 768 | |
| 769 | Error collectVTableStrings(ArrayRef<GlobalVariable *> VTables, |
| 770 | std::string &Result, bool DoCompression) { |
| 771 | std::vector<std::string> VTableNameStrs; |
| 772 | for (auto *VTable : VTables) |
| 773 | VTableNameStrs.push_back(x: getPGOName(V: *VTable)); |
| 774 | return collectGlobalObjectNameStrings( |
| 775 | NameStrs: VTableNameStrs, DoCompression: compression::zlib::isAvailable() && DoCompression, |
| 776 | Result); |
| 777 | } |
| 778 | |
| 779 | void InstrProfRecord::accumulateCounts(CountSumOrPercent &Sum) const { |
| 780 | uint64_t FuncSum = 0; |
| 781 | Sum.NumEntries += Counts.size(); |
| 782 | for (uint64_t Count : Counts) |
| 783 | FuncSum += Count; |
| 784 | Sum.CountSum += FuncSum; |
| 785 | |
| 786 | for (uint32_t VK = IPVK_First; VK <= IPVK_Last; ++VK) { |
| 787 | uint64_t KindSum = 0; |
| 788 | uint32_t NumValueSites = getNumValueSites(ValueKind: VK); |
| 789 | for (size_t I = 0; I < NumValueSites; ++I) { |
| 790 | for (const auto &V : getValueArrayForSite(ValueKind: VK, Site: I)) |
| 791 | KindSum += V.Count; |
| 792 | } |
| 793 | Sum.ValueCounts[VK] += KindSum; |
| 794 | } |
| 795 | } |
| 796 | |
| 797 | void InstrProfValueSiteRecord::overlap(InstrProfValueSiteRecord &Input, |
| 798 | uint32_t ValueKind, |
| 799 | OverlapStats &Overlap, |
| 800 | OverlapStats &FuncLevelOverlap) { |
| 801 | this->sortByTargetValues(); |
| 802 | Input.sortByTargetValues(); |
| 803 | double Score = 0.0f, FuncLevelScore = 0.0f; |
| 804 | auto I = ValueData.begin(); |
| 805 | auto IE = ValueData.end(); |
| 806 | auto J = Input.ValueData.begin(); |
| 807 | auto JE = Input.ValueData.end(); |
| 808 | while (I != IE && J != JE) { |
| 809 | if (I->Value == J->Value) { |
| 810 | Score += OverlapStats::score(Val1: I->Count, Val2: J->Count, |
| 811 | Sum1: Overlap.Base.ValueCounts[ValueKind], |
| 812 | Sum2: Overlap.Test.ValueCounts[ValueKind]); |
| 813 | FuncLevelScore += OverlapStats::score( |
| 814 | Val1: I->Count, Val2: J->Count, Sum1: FuncLevelOverlap.Base.ValueCounts[ValueKind], |
| 815 | Sum2: FuncLevelOverlap.Test.ValueCounts[ValueKind]); |
| 816 | ++I; |
| 817 | } else if (I->Value < J->Value) { |
| 818 | ++I; |
| 819 | continue; |
| 820 | } |
| 821 | ++J; |
| 822 | } |
| 823 | Overlap.Overlap.ValueCounts[ValueKind] += Score; |
| 824 | FuncLevelOverlap.Overlap.ValueCounts[ValueKind] += FuncLevelScore; |
| 825 | } |
| 826 | |
| 827 | // Return false on mismatch. |
| 828 | void InstrProfRecord::overlapValueProfData(uint32_t ValueKind, |
| 829 | InstrProfRecord &Other, |
| 830 | OverlapStats &Overlap, |
| 831 | OverlapStats &FuncLevelOverlap) { |
| 832 | uint32_t ThisNumValueSites = getNumValueSites(ValueKind); |
| 833 | assert(ThisNumValueSites == Other.getNumValueSites(ValueKind)); |
| 834 | if (!ThisNumValueSites) |
| 835 | return; |
| 836 | |
| 837 | std::vector<InstrProfValueSiteRecord> &ThisSiteRecords = |
| 838 | getOrCreateValueSitesForKind(ValueKind); |
| 839 | MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords = |
| 840 | Other.getValueSitesForKind(ValueKind); |
| 841 | for (uint32_t I = 0; I < ThisNumValueSites; I++) |
| 842 | ThisSiteRecords[I].overlap(Input&: OtherSiteRecords[I], ValueKind, Overlap, |
| 843 | FuncLevelOverlap); |
| 844 | } |
| 845 | |
| 846 | void InstrProfRecord::overlap(InstrProfRecord &Other, OverlapStats &Overlap, |
| 847 | OverlapStats &FuncLevelOverlap, |
| 848 | uint64_t ValueCutoff) { |
| 849 | // FuncLevel CountSum for other should already computed and nonzero. |
| 850 | assert(FuncLevelOverlap.Test.CountSum >= 1.0f); |
| 851 | accumulateCounts(Sum&: FuncLevelOverlap.Base); |
| 852 | bool Mismatch = (Counts.size() != Other.Counts.size()); |
| 853 | |
| 854 | // Check if the value profiles mismatch. |
| 855 | if (!Mismatch) { |
| 856 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) { |
| 857 | uint32_t ThisNumValueSites = getNumValueSites(ValueKind: Kind); |
| 858 | uint32_t OtherNumValueSites = Other.getNumValueSites(ValueKind: Kind); |
| 859 | if (ThisNumValueSites != OtherNumValueSites) { |
| 860 | Mismatch = true; |
| 861 | break; |
| 862 | } |
| 863 | } |
| 864 | } |
| 865 | if (Mismatch) { |
| 866 | Overlap.addOneMismatch(MismatchFunc: FuncLevelOverlap.Test); |
| 867 | return; |
| 868 | } |
| 869 | |
| 870 | // Compute overlap for value counts. |
| 871 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) |
| 872 | overlapValueProfData(ValueKind: Kind, Other, Overlap, FuncLevelOverlap); |
| 873 | |
| 874 | double Score = 0.0; |
| 875 | uint64_t MaxCount = 0; |
| 876 | // Compute overlap for edge counts. |
| 877 | for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) { |
| 878 | Score += OverlapStats::score(Val1: Counts[I], Val2: Other.Counts[I], |
| 879 | Sum1: Overlap.Base.CountSum, Sum2: Overlap.Test.CountSum); |
| 880 | MaxCount = std::max(a: Other.Counts[I], b: MaxCount); |
| 881 | } |
| 882 | Overlap.Overlap.CountSum += Score; |
| 883 | Overlap.Overlap.NumEntries += 1; |
| 884 | |
| 885 | if (MaxCount >= ValueCutoff) { |
| 886 | double FuncScore = 0.0; |
| 887 | for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) |
| 888 | FuncScore += OverlapStats::score(Val1: Counts[I], Val2: Other.Counts[I], |
| 889 | Sum1: FuncLevelOverlap.Base.CountSum, |
| 890 | Sum2: FuncLevelOverlap.Test.CountSum); |
| 891 | FuncLevelOverlap.Overlap.CountSum = FuncScore; |
| 892 | FuncLevelOverlap.Overlap.NumEntries = Other.Counts.size(); |
| 893 | FuncLevelOverlap.Valid = true; |
| 894 | } |
| 895 | } |
| 896 | |
| 897 | void InstrProfValueSiteRecord::merge(InstrProfValueSiteRecord &Input, |
| 898 | uint64_t Weight, |
| 899 | function_ref<void(instrprof_error)> Warn) { |
| 900 | this->sortByTargetValues(); |
| 901 | Input.sortByTargetValues(); |
| 902 | auto I = ValueData.begin(); |
| 903 | auto IE = ValueData.end(); |
| 904 | std::vector<InstrProfValueData> Merged; |
| 905 | Merged.reserve(n: std::max(a: ValueData.size(), b: Input.ValueData.size())); |
| 906 | for (const InstrProfValueData &J : Input.ValueData) { |
| 907 | while (I != IE && I->Value < J.Value) { |
| 908 | Merged.push_back(x: *I); |
| 909 | ++I; |
| 910 | } |
| 911 | if (I != IE && I->Value == J.Value) { |
| 912 | bool Overflowed; |
| 913 | I->Count = SaturatingMultiplyAdd(X: J.Count, Y: Weight, A: I->Count, ResultOverflowed: &Overflowed); |
| 914 | if (Overflowed) |
| 915 | Warn(instrprof_error::counter_overflow); |
| 916 | Merged.push_back(x: *I); |
| 917 | ++I; |
| 918 | continue; |
| 919 | } |
| 920 | Merged.push_back(x: J); |
| 921 | } |
| 922 | Merged.insert(position: Merged.end(), first: I, last: IE); |
| 923 | ValueData = std::move(Merged); |
| 924 | } |
| 925 | |
| 926 | void InstrProfValueSiteRecord::scale(uint64_t N, uint64_t D, |
| 927 | function_ref<void(instrprof_error)> Warn) { |
| 928 | for (InstrProfValueData &I : ValueData) { |
| 929 | bool Overflowed; |
| 930 | I.Count = SaturatingMultiply(X: I.Count, Y: N, ResultOverflowed: &Overflowed) / D; |
| 931 | if (Overflowed) |
| 932 | Warn(instrprof_error::counter_overflow); |
| 933 | } |
| 934 | } |
| 935 | |
| 936 | // Merge Value Profile data from Src record to this record for ValueKind. |
| 937 | // Scale merged value counts by \p Weight. |
| 938 | void InstrProfRecord::mergeValueProfData( |
| 939 | uint32_t ValueKind, InstrProfRecord &Src, uint64_t Weight, |
| 940 | function_ref<void(instrprof_error)> Warn) { |
| 941 | uint32_t ThisNumValueSites = getNumValueSites(ValueKind); |
| 942 | uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind); |
| 943 | if (ThisNumValueSites != OtherNumValueSites) { |
| 944 | Warn(instrprof_error::value_site_count_mismatch); |
| 945 | return; |
| 946 | } |
| 947 | if (!ThisNumValueSites) |
| 948 | return; |
| 949 | std::vector<InstrProfValueSiteRecord> &ThisSiteRecords = |
| 950 | getOrCreateValueSitesForKind(ValueKind); |
| 951 | MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords = |
| 952 | Src.getValueSitesForKind(ValueKind); |
| 953 | for (uint32_t I = 0; I < ThisNumValueSites; I++) |
| 954 | ThisSiteRecords[I].merge(Input&: OtherSiteRecords[I], Weight, Warn); |
| 955 | } |
| 956 | |
| 957 | void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight, |
| 958 | function_ref<void(instrprof_error)> Warn) { |
| 959 | // If the number of counters doesn't match we either have bad data |
| 960 | // or a hash collision. |
| 961 | if (Counts.size() != Other.Counts.size()) { |
| 962 | Warn(instrprof_error::count_mismatch); |
| 963 | return; |
| 964 | } |
| 965 | |
| 966 | // Special handling of the first count as the PseudoCount. |
| 967 | CountPseudoKind OtherKind = Other.getCountPseudoKind(); |
| 968 | CountPseudoKind ThisKind = getCountPseudoKind(); |
| 969 | if (OtherKind != NotPseudo || ThisKind != NotPseudo) { |
| 970 | // We don't allow the merge of a profile with pseudo counts and |
| 971 | // a normal profile (i.e. without pesudo counts). |
| 972 | // Profile supplimenation should be done after the profile merge. |
| 973 | if (OtherKind == NotPseudo || ThisKind == NotPseudo) { |
| 974 | Warn(instrprof_error::count_mismatch); |
| 975 | return; |
| 976 | } |
| 977 | if (OtherKind == PseudoHot || ThisKind == PseudoHot) |
| 978 | setPseudoCount(PseudoHot); |
| 979 | else |
| 980 | setPseudoCount(PseudoWarm); |
| 981 | return; |
| 982 | } |
| 983 | |
| 984 | for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) { |
| 985 | bool Overflowed; |
| 986 | uint64_t Value = |
| 987 | SaturatingMultiplyAdd(X: Other.Counts[I], Y: Weight, A: Counts[I], ResultOverflowed: &Overflowed); |
| 988 | if (Value > getInstrMaxCountValue()) { |
| 989 | Value = getInstrMaxCountValue(); |
| 990 | Overflowed = true; |
| 991 | } |
| 992 | Counts[I] = Value; |
| 993 | if (Overflowed) |
| 994 | Warn(instrprof_error::counter_overflow); |
| 995 | } |
| 996 | |
| 997 | // If the number of bitmap bytes doesn't match we either have bad data |
| 998 | // or a hash collision. |
| 999 | if (BitmapBytes.size() != Other.BitmapBytes.size()) { |
| 1000 | Warn(instrprof_error::bitmap_mismatch); |
| 1001 | return; |
| 1002 | } |
| 1003 | |
| 1004 | // Bitmap bytes are merged by simply ORing them together. |
| 1005 | for (size_t I = 0, E = Other.BitmapBytes.size(); I < E; ++I) { |
| 1006 | BitmapBytes[I] = Other.BitmapBytes[I] | BitmapBytes[I]; |
| 1007 | } |
| 1008 | |
| 1009 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) |
| 1010 | mergeValueProfData(ValueKind: Kind, Src&: Other, Weight, Warn); |
| 1011 | } |
| 1012 | |
| 1013 | void InstrProfRecord::scaleValueProfData( |
| 1014 | uint32_t ValueKind, uint64_t N, uint64_t D, |
| 1015 | function_ref<void(instrprof_error)> Warn) { |
| 1016 | for (auto &R : getValueSitesForKind(ValueKind)) |
| 1017 | R.scale(N, D, Warn); |
| 1018 | } |
| 1019 | |
| 1020 | void InstrProfRecord::scale(uint64_t N, uint64_t D, |
| 1021 | function_ref<void(instrprof_error)> Warn) { |
| 1022 | assert(D != 0 && "D cannot be 0" ); |
| 1023 | for (auto &Count : this->Counts) { |
| 1024 | bool Overflowed; |
| 1025 | Count = SaturatingMultiply(X: Count, Y: N, ResultOverflowed: &Overflowed) / D; |
| 1026 | if (Count > getInstrMaxCountValue()) { |
| 1027 | Count = getInstrMaxCountValue(); |
| 1028 | Overflowed = true; |
| 1029 | } |
| 1030 | if (Overflowed) |
| 1031 | Warn(instrprof_error::counter_overflow); |
| 1032 | } |
| 1033 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) |
| 1034 | scaleValueProfData(ValueKind: Kind, N, D, Warn); |
| 1035 | } |
| 1036 | |
| 1037 | // Map indirect call target name hash to name string. |
| 1038 | uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind, |
| 1039 | InstrProfSymtab *SymTab) { |
| 1040 | if (!SymTab) |
| 1041 | return Value; |
| 1042 | |
| 1043 | if (ValueKind == IPVK_IndirectCallTarget) |
| 1044 | return SymTab->getFunctionHashFromAddress(Address: Value); |
| 1045 | |
| 1046 | if (ValueKind == IPVK_VTableTarget) |
| 1047 | return SymTab->getVTableHashFromAddress(Address: Value); |
| 1048 | |
| 1049 | return Value; |
| 1050 | } |
| 1051 | |
| 1052 | void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site, |
| 1053 | ArrayRef<InstrProfValueData> VData, |
| 1054 | InstrProfSymtab *ValueMap) { |
| 1055 | // Remap values. |
| 1056 | std::vector<InstrProfValueData> RemappedVD; |
| 1057 | RemappedVD.reserve(n: VData.size()); |
| 1058 | for (const auto &V : VData) { |
| 1059 | uint64_t NewValue = remapValue(Value: V.Value, ValueKind, SymTab: ValueMap); |
| 1060 | RemappedVD.push_back(x: {.Value: NewValue, .Count: V.Count}); |
| 1061 | } |
| 1062 | |
| 1063 | std::vector<InstrProfValueSiteRecord> &ValueSites = |
| 1064 | getOrCreateValueSitesForKind(ValueKind); |
| 1065 | assert(ValueSites.size() == Site); |
| 1066 | |
| 1067 | // Add a new value site with remapped value profiling data. |
| 1068 | ValueSites.emplace_back(args: std::move(RemappedVD)); |
| 1069 | } |
| 1070 | |
| 1071 | void TemporalProfTraceTy::createBPFunctionNodes( |
| 1072 | ArrayRef<TemporalProfTraceTy> Traces, std::vector<BPFunctionNode> &Nodes, |
| 1073 | bool RemoveOutlierUNs) { |
| 1074 | using IDT = BPFunctionNode::IDT; |
| 1075 | using UtilityNodeT = BPFunctionNode::UtilityNodeT; |
| 1076 | UtilityNodeT MaxUN = 0; |
| 1077 | DenseMap<IDT, size_t> IdToFirstTimestamp; |
| 1078 | DenseMap<IDT, UtilityNodeT> IdToFirstUN; |
| 1079 | DenseMap<IDT, SmallVector<UtilityNodeT>> IdToUNs; |
| 1080 | // TODO: We need to use the Trace.Weight field to give more weight to more |
| 1081 | // important utilities |
| 1082 | for (auto &Trace : Traces) { |
| 1083 | size_t CutoffTimestamp = 1; |
| 1084 | for (size_t Timestamp = 0; Timestamp < Trace.FunctionNameRefs.size(); |
| 1085 | Timestamp++) { |
| 1086 | IDT Id = Trace.FunctionNameRefs[Timestamp]; |
| 1087 | auto [It, WasInserted] = IdToFirstTimestamp.try_emplace(Key: Id, Args&: Timestamp); |
| 1088 | if (!WasInserted) |
| 1089 | It->getSecond() = std::min<size_t>(a: It->getSecond(), b: Timestamp); |
| 1090 | if (Timestamp >= CutoffTimestamp) { |
| 1091 | ++MaxUN; |
| 1092 | CutoffTimestamp = 2 * Timestamp; |
| 1093 | } |
| 1094 | IdToFirstUN.try_emplace(Key: Id, Args&: MaxUN); |
| 1095 | } |
| 1096 | for (auto &[Id, FirstUN] : IdToFirstUN) |
| 1097 | for (auto UN = FirstUN; UN <= MaxUN; ++UN) |
| 1098 | IdToUNs[Id].push_back(Elt: UN); |
| 1099 | ++MaxUN; |
| 1100 | IdToFirstUN.clear(); |
| 1101 | } |
| 1102 | |
| 1103 | if (RemoveOutlierUNs) { |
| 1104 | DenseMap<UtilityNodeT, unsigned> UNFrequency; |
| 1105 | for (auto &[Id, UNs] : IdToUNs) |
| 1106 | for (auto &UN : UNs) |
| 1107 | ++UNFrequency[UN]; |
| 1108 | // Filter out utility nodes that are too infrequent or too prevalent to make |
| 1109 | // BalancedPartitioning more effective. |
| 1110 | for (auto &[Id, UNs] : IdToUNs) |
| 1111 | llvm::erase_if(C&: UNs, P: [&](auto &UN) { |
| 1112 | unsigned Freq = UNFrequency[UN]; |
| 1113 | return Freq <= 1 || 2 * Freq > IdToUNs.size(); |
| 1114 | }); |
| 1115 | } |
| 1116 | |
| 1117 | for (auto &[Id, UNs] : IdToUNs) |
| 1118 | Nodes.emplace_back(args&: Id, args&: UNs); |
| 1119 | |
| 1120 | // Since BalancedPartitioning is sensitive to the initial order, we explicitly |
| 1121 | // order nodes by their earliest timestamp. |
| 1122 | llvm::sort(C&: Nodes, Comp: [&](auto &L, auto &R) { |
| 1123 | return std::make_pair(IdToFirstTimestamp[L.Id], L.Id) < |
| 1124 | std::make_pair(IdToFirstTimestamp[R.Id], R.Id); |
| 1125 | }); |
| 1126 | } |
| 1127 | |
| 1128 | #define INSTR_PROF_COMMON_API_IMPL |
| 1129 | #include "llvm/ProfileData/InstrProfData.inc" |
| 1130 | |
| 1131 | /*! |
| 1132 | * ValueProfRecordClosure Interface implementation for InstrProfRecord |
| 1133 | * class. These C wrappers are used as adaptors so that C++ code can be |
| 1134 | * invoked as callbacks. |
| 1135 | */ |
| 1136 | uint32_t getNumValueKindsInstrProf(const void *Record) { |
| 1137 | return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds(); |
| 1138 | } |
| 1139 | |
| 1140 | uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) { |
| 1141 | return reinterpret_cast<const InstrProfRecord *>(Record) |
| 1142 | ->getNumValueSites(ValueKind: VKind); |
| 1143 | } |
| 1144 | |
| 1145 | uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) { |
| 1146 | return reinterpret_cast<const InstrProfRecord *>(Record) |
| 1147 | ->getNumValueData(ValueKind: VKind); |
| 1148 | } |
| 1149 | |
| 1150 | uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK, |
| 1151 | uint32_t S) { |
| 1152 | const auto *IPR = reinterpret_cast<const InstrProfRecord *>(R); |
| 1153 | return IPR->getValueArrayForSite(ValueKind: VK, Site: S).size(); |
| 1154 | } |
| 1155 | |
| 1156 | void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst, |
| 1157 | uint32_t K, uint32_t S) { |
| 1158 | const auto *IPR = reinterpret_cast<const InstrProfRecord *>(R); |
| 1159 | llvm::copy(Range: IPR->getValueArrayForSite(ValueKind: K, Site: S), Out: Dst); |
| 1160 | } |
| 1161 | |
| 1162 | ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) { |
| 1163 | ValueProfData *VD = |
| 1164 | (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData()); |
| 1165 | memset(s: VD, c: 0, n: TotalSizeInBytes); |
| 1166 | return VD; |
| 1167 | } |
| 1168 | |
| 1169 | static ValueProfRecordClosure InstrProfRecordClosure = { |
| 1170 | .Record: nullptr, |
| 1171 | .GetNumValueKinds: getNumValueKindsInstrProf, |
| 1172 | .GetNumValueSites: getNumValueSitesInstrProf, |
| 1173 | .GetNumValueData: getNumValueDataInstrProf, |
| 1174 | .GetNumValueDataForSite: getNumValueDataForSiteInstrProf, |
| 1175 | .RemapValueData: nullptr, |
| 1176 | .GetValueForSite: getValueForSiteInstrProf, |
| 1177 | .AllocValueProfData: allocValueProfDataInstrProf}; |
| 1178 | |
| 1179 | // Wrapper implementation using the closure mechanism. |
| 1180 | uint32_t ValueProfData::getSize(const InstrProfRecord &Record) { |
| 1181 | auto Closure = InstrProfRecordClosure; |
| 1182 | Closure.Record = &Record; |
| 1183 | return getValueProfDataSize(Closure: &Closure); |
| 1184 | } |
| 1185 | |
| 1186 | // Wrapper implementation using the closure mechanism. |
| 1187 | std::unique_ptr<ValueProfData> |
| 1188 | ValueProfData::serializeFrom(const InstrProfRecord &Record) { |
| 1189 | InstrProfRecordClosure.Record = &Record; |
| 1190 | |
| 1191 | std::unique_ptr<ValueProfData> VPD( |
| 1192 | serializeValueProfDataFrom(Closure: &InstrProfRecordClosure, DstData: nullptr)); |
| 1193 | return VPD; |
| 1194 | } |
| 1195 | |
| 1196 | void ValueProfRecord::deserializeTo(InstrProfRecord &Record, |
| 1197 | InstrProfSymtab *SymTab) { |
| 1198 | Record.reserveSites(ValueKind: Kind, NumValueSites); |
| 1199 | |
| 1200 | InstrProfValueData *ValueData = getValueProfRecordValueData(This: this); |
| 1201 | for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) { |
| 1202 | uint8_t ValueDataCount = this->SiteCountArray[VSite]; |
| 1203 | ArrayRef<InstrProfValueData> VDs(ValueData, ValueDataCount); |
| 1204 | Record.addValueData(ValueKind: Kind, Site: VSite, VData: VDs, ValueMap: SymTab); |
| 1205 | ValueData += ValueDataCount; |
| 1206 | } |
| 1207 | } |
| 1208 | |
| 1209 | // For writing/serializing, Old is the host endianness, and New is |
| 1210 | // byte order intended on disk. For Reading/deserialization, Old |
| 1211 | // is the on-disk source endianness, and New is the host endianness. |
| 1212 | void ValueProfRecord::swapBytes(llvm::endianness Old, llvm::endianness New) { |
| 1213 | using namespace support; |
| 1214 | |
| 1215 | if (Old == New) |
| 1216 | return; |
| 1217 | |
| 1218 | if (llvm::endianness::native != Old) { |
| 1219 | sys::swapByteOrder<uint32_t>(Value&: NumValueSites); |
| 1220 | sys::swapByteOrder<uint32_t>(Value&: Kind); |
| 1221 | } |
| 1222 | uint32_t ND = getValueProfRecordNumValueData(This: this); |
| 1223 | InstrProfValueData *VD = getValueProfRecordValueData(This: this); |
| 1224 | |
| 1225 | // No need to swap byte array: SiteCountArrray. |
| 1226 | for (uint32_t I = 0; I < ND; I++) { |
| 1227 | sys::swapByteOrder<uint64_t>(Value&: VD[I].Value); |
| 1228 | sys::swapByteOrder<uint64_t>(Value&: VD[I].Count); |
| 1229 | } |
| 1230 | if (llvm::endianness::native == Old) { |
| 1231 | sys::swapByteOrder<uint32_t>(Value&: NumValueSites); |
| 1232 | sys::swapByteOrder<uint32_t>(Value&: Kind); |
| 1233 | } |
| 1234 | } |
| 1235 | |
| 1236 | void ValueProfData::deserializeTo(InstrProfRecord &Record, |
| 1237 | InstrProfSymtab *SymTab) { |
| 1238 | if (NumValueKinds == 0) |
| 1239 | return; |
| 1240 | |
| 1241 | ValueProfRecord *VR = getFirstValueProfRecord(This: this); |
| 1242 | for (uint32_t K = 0; K < NumValueKinds; K++) { |
| 1243 | VR->deserializeTo(Record, SymTab); |
| 1244 | VR = getValueProfRecordNext(This: VR); |
| 1245 | } |
| 1246 | } |
| 1247 | |
| 1248 | static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) { |
| 1249 | return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize)) |
| 1250 | ValueProfData()); |
| 1251 | } |
| 1252 | |
| 1253 | Error ValueProfData::checkIntegrity() { |
| 1254 | if (NumValueKinds > IPVK_Last + 1) |
| 1255 | return make_error<InstrProfError>( |
| 1256 | Args: instrprof_error::malformed, Args: "number of value profile kinds is invalid" ); |
| 1257 | // Total size needs to be multiple of quadword size. |
| 1258 | if (TotalSize % sizeof(uint64_t)) |
| 1259 | return make_error<InstrProfError>( |
| 1260 | Args: instrprof_error::malformed, Args: "total size is not multiples of quardword" ); |
| 1261 | |
| 1262 | ValueProfRecord *VR = getFirstValueProfRecord(This: this); |
| 1263 | for (uint32_t K = 0; K < this->NumValueKinds; K++) { |
| 1264 | if (VR->Kind > IPVK_Last) |
| 1265 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
| 1266 | Args: "value kind is invalid" ); |
| 1267 | VR = getValueProfRecordNext(This: VR); |
| 1268 | if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize) |
| 1269 | return make_error<InstrProfError>( |
| 1270 | Args: instrprof_error::malformed, |
| 1271 | Args: "value profile address is greater than total size" ); |
| 1272 | } |
| 1273 | return Error::success(); |
| 1274 | } |
| 1275 | |
| 1276 | Expected<std::unique_ptr<ValueProfData>> |
| 1277 | ValueProfData::getValueProfData(const unsigned char *D, |
| 1278 | const unsigned char *const BufferEnd, |
| 1279 | llvm::endianness Endianness) { |
| 1280 | using namespace support; |
| 1281 | |
| 1282 | if (D + sizeof(ValueProfData) > BufferEnd) |
| 1283 | return make_error<InstrProfError>(Args: instrprof_error::truncated); |
| 1284 | |
| 1285 | const unsigned char * = D; |
| 1286 | uint32_t TotalSize = endian::readNext<uint32_t>(memory&: Header, endian: Endianness); |
| 1287 | |
| 1288 | if (D + TotalSize > BufferEnd) |
| 1289 | return make_error<InstrProfError>(Args: instrprof_error::too_large); |
| 1290 | |
| 1291 | std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize); |
| 1292 | memcpy(dest: VPD.get(), src: D, n: TotalSize); |
| 1293 | // Byte swap. |
| 1294 | VPD->swapBytesToHost(Endianness); |
| 1295 | |
| 1296 | Error E = VPD->checkIntegrity(); |
| 1297 | if (E) |
| 1298 | return std::move(E); |
| 1299 | |
| 1300 | return std::move(VPD); |
| 1301 | } |
| 1302 | |
| 1303 | void ValueProfData::swapBytesToHost(llvm::endianness Endianness) { |
| 1304 | using namespace support; |
| 1305 | |
| 1306 | if (Endianness == llvm::endianness::native) |
| 1307 | return; |
| 1308 | |
| 1309 | sys::swapByteOrder<uint32_t>(Value&: TotalSize); |
| 1310 | sys::swapByteOrder<uint32_t>(Value&: NumValueKinds); |
| 1311 | |
| 1312 | ValueProfRecord *VR = getFirstValueProfRecord(This: this); |
| 1313 | for (uint32_t K = 0; K < NumValueKinds; K++) { |
| 1314 | VR->swapBytes(Old: Endianness, New: llvm::endianness::native); |
| 1315 | VR = getValueProfRecordNext(This: VR); |
| 1316 | } |
| 1317 | } |
| 1318 | |
| 1319 | void ValueProfData::swapBytesFromHost(llvm::endianness Endianness) { |
| 1320 | using namespace support; |
| 1321 | |
| 1322 | if (Endianness == llvm::endianness::native) |
| 1323 | return; |
| 1324 | |
| 1325 | ValueProfRecord *VR = getFirstValueProfRecord(This: this); |
| 1326 | for (uint32_t K = 0; K < NumValueKinds; K++) { |
| 1327 | ValueProfRecord *NVR = getValueProfRecordNext(This: VR); |
| 1328 | VR->swapBytes(Old: llvm::endianness::native, New: Endianness); |
| 1329 | VR = NVR; |
| 1330 | } |
| 1331 | sys::swapByteOrder<uint32_t>(Value&: TotalSize); |
| 1332 | sys::swapByteOrder<uint32_t>(Value&: NumValueKinds); |
| 1333 | } |
| 1334 | |
| 1335 | void annotateValueSite(Module &M, Instruction &Inst, |
| 1336 | const InstrProfRecord &InstrProfR, |
| 1337 | InstrProfValueKind ValueKind, uint32_t SiteIdx, |
| 1338 | uint32_t MaxMDCount) { |
| 1339 | auto VDs = InstrProfR.getValueArrayForSite(ValueKind, Site: SiteIdx); |
| 1340 | if (VDs.empty()) |
| 1341 | return; |
| 1342 | uint64_t Sum = 0; |
| 1343 | for (const InstrProfValueData &V : VDs) |
| 1344 | Sum = SaturatingAdd(X: Sum, Y: V.Count); |
| 1345 | annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount); |
| 1346 | } |
| 1347 | |
| 1348 | void annotateValueSite(Module &M, Instruction &Inst, |
| 1349 | ArrayRef<InstrProfValueData> VDs, |
| 1350 | uint64_t Sum, InstrProfValueKind ValueKind, |
| 1351 | uint32_t MaxMDCount) { |
| 1352 | if (VDs.empty()) |
| 1353 | return; |
| 1354 | LLVMContext &Ctx = M.getContext(); |
| 1355 | MDBuilder MDHelper(Ctx); |
| 1356 | SmallVector<Metadata *, 3> Vals; |
| 1357 | // Tag |
| 1358 | Vals.push_back(Elt: MDHelper.createString(Str: MDProfLabels::ValueProfile)); |
| 1359 | // Value Kind |
| 1360 | Vals.push_back(Elt: MDHelper.createConstant( |
| 1361 | C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: ValueKind))); |
| 1362 | // Total Count |
| 1363 | Vals.push_back( |
| 1364 | Elt: MDHelper.createConstant(C: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: Sum))); |
| 1365 | |
| 1366 | // Value Profile Data |
| 1367 | uint32_t MDCount = MaxMDCount; |
| 1368 | for (const auto &VD : VDs) { |
| 1369 | Vals.push_back(Elt: MDHelper.createConstant( |
| 1370 | C: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: VD.Value))); |
| 1371 | Vals.push_back(Elt: MDHelper.createConstant( |
| 1372 | C: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: VD.Count))); |
| 1373 | if (--MDCount == 0) |
| 1374 | break; |
| 1375 | } |
| 1376 | Inst.setMetadata(KindID: LLVMContext::MD_prof, Node: MDNode::get(Context&: Ctx, MDs: Vals)); |
| 1377 | } |
| 1378 | |
| 1379 | MDNode *mayHaveValueProfileOfKind(const Instruction &Inst, |
| 1380 | InstrProfValueKind ValueKind) { |
| 1381 | MDNode *MD = Inst.getMetadata(KindID: LLVMContext::MD_prof); |
| 1382 | if (!MD) |
| 1383 | return nullptr; |
| 1384 | |
| 1385 | if (MD->getNumOperands() < 5) |
| 1386 | return nullptr; |
| 1387 | |
| 1388 | MDString *Tag = cast<MDString>(Val: MD->getOperand(I: 0)); |
| 1389 | if (!Tag || Tag->getString() != MDProfLabels::ValueProfile) |
| 1390 | return nullptr; |
| 1391 | |
| 1392 | // Now check kind: |
| 1393 | ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD: MD->getOperand(I: 1)); |
| 1394 | if (!KindInt) |
| 1395 | return nullptr; |
| 1396 | if (KindInt->getZExtValue() != ValueKind) |
| 1397 | return nullptr; |
| 1398 | |
| 1399 | return MD; |
| 1400 | } |
| 1401 | |
| 1402 | SmallVector<InstrProfValueData, 4> |
| 1403 | getValueProfDataFromInst(const Instruction &Inst, InstrProfValueKind ValueKind, |
| 1404 | uint32_t MaxNumValueData, uint64_t &TotalC, |
| 1405 | bool GetNoICPValue) { |
| 1406 | // Four inline elements seem to work well in practice. With MaxNumValueData, |
| 1407 | // this array won't grow very big anyway. |
| 1408 | SmallVector<InstrProfValueData, 4> ValueData; |
| 1409 | MDNode *MD = mayHaveValueProfileOfKind(Inst, ValueKind); |
| 1410 | if (!MD) |
| 1411 | return ValueData; |
| 1412 | const unsigned NOps = MD->getNumOperands(); |
| 1413 | // Get total count |
| 1414 | ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD: MD->getOperand(I: 2)); |
| 1415 | if (!TotalCInt) |
| 1416 | return ValueData; |
| 1417 | TotalC = TotalCInt->getZExtValue(); |
| 1418 | |
| 1419 | ValueData.reserve(N: (NOps - 3) / 2); |
| 1420 | for (unsigned I = 3; I < NOps; I += 2) { |
| 1421 | if (ValueData.size() >= MaxNumValueData) |
| 1422 | break; |
| 1423 | ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD: MD->getOperand(I)); |
| 1424 | ConstantInt *Count = |
| 1425 | mdconst::dyn_extract<ConstantInt>(MD: MD->getOperand(I: I + 1)); |
| 1426 | if (!Value || !Count) { |
| 1427 | ValueData.clear(); |
| 1428 | return ValueData; |
| 1429 | } |
| 1430 | uint64_t CntValue = Count->getZExtValue(); |
| 1431 | if (!GetNoICPValue && (CntValue == NOMORE_ICP_MAGICNUM)) |
| 1432 | continue; |
| 1433 | InstrProfValueData V; |
| 1434 | V.Value = Value->getZExtValue(); |
| 1435 | V.Count = CntValue; |
| 1436 | ValueData.push_back(Elt: V); |
| 1437 | } |
| 1438 | return ValueData; |
| 1439 | } |
| 1440 | |
| 1441 | MDNode *getPGOFuncNameMetadata(const Function &F) { |
| 1442 | return F.getMetadata(Kind: getPGOFuncNameMetadataName()); |
| 1443 | } |
| 1444 | |
| 1445 | static void createPGONameMetadata(GlobalObject &GO, StringRef MetadataName, |
| 1446 | StringRef PGOName) { |
| 1447 | // Only for internal linkage functions or global variables. The name is not |
| 1448 | // the same as PGO name for these global objects. |
| 1449 | if (GO.getName() == PGOName) |
| 1450 | return; |
| 1451 | |
| 1452 | // Don't create duplicated metadata. |
| 1453 | if (GO.getMetadata(Kind: MetadataName)) |
| 1454 | return; |
| 1455 | |
| 1456 | LLVMContext &C = GO.getContext(); |
| 1457 | MDNode *N = MDNode::get(Context&: C, MDs: MDString::get(Context&: C, Str: PGOName)); |
| 1458 | GO.setMetadata(Kind: MetadataName, Node: N); |
| 1459 | } |
| 1460 | |
| 1461 | void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) { |
| 1462 | return createPGONameMetadata(GO&: F, MetadataName: getPGOFuncNameMetadataName(), PGOName: PGOFuncName); |
| 1463 | } |
| 1464 | |
| 1465 | void createPGONameMetadata(GlobalObject &GO, StringRef PGOName) { |
| 1466 | return createPGONameMetadata(GO, MetadataName: getPGONameMetadataName(), PGOName); |
| 1467 | } |
| 1468 | |
| 1469 | bool needsComdatForCounter(const GlobalObject &GO, const Module &M) { |
| 1470 | if (GO.hasComdat()) |
| 1471 | return true; |
| 1472 | |
| 1473 | if (!M.getTargetTriple().supportsCOMDAT()) |
| 1474 | return false; |
| 1475 | |
| 1476 | // See createPGOFuncNameVar for more details. To avoid link errors, profile |
| 1477 | // counters for function with available_externally linkage needs to be changed |
| 1478 | // to linkonce linkage. On ELF based systems, this leads to weak symbols to be |
| 1479 | // created. Without using comdat, duplicate entries won't be removed by the |
| 1480 | // linker leading to increased data segement size and raw profile size. Even |
| 1481 | // worse, since the referenced counter from profile per-function data object |
| 1482 | // will be resolved to the common strong definition, the profile counts for |
| 1483 | // available_externally functions will end up being duplicated in raw profile |
| 1484 | // data. This can result in distorted profile as the counts of those dups |
| 1485 | // will be accumulated by the profile merger. |
| 1486 | GlobalValue::LinkageTypes Linkage = GO.getLinkage(); |
| 1487 | if (Linkage != GlobalValue::ExternalWeakLinkage && |
| 1488 | Linkage != GlobalValue::AvailableExternallyLinkage) |
| 1489 | return false; |
| 1490 | |
| 1491 | return true; |
| 1492 | } |
| 1493 | |
| 1494 | // Check if INSTR_PROF_RAW_VERSION_VAR is defined. |
| 1495 | bool isIRPGOFlagSet(const Module *M) { |
| 1496 | const GlobalVariable *IRInstrVar = |
| 1497 | M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR)); |
| 1498 | if (!IRInstrVar || IRInstrVar->hasLocalLinkage()) |
| 1499 | return false; |
| 1500 | |
| 1501 | // For CSPGO+LTO, this variable might be marked as non-prevailing and we only |
| 1502 | // have the decl. |
| 1503 | if (IRInstrVar->isDeclaration()) |
| 1504 | return true; |
| 1505 | |
| 1506 | // Check if the flag is set. |
| 1507 | if (!IRInstrVar->hasInitializer()) |
| 1508 | return false; |
| 1509 | |
| 1510 | auto *InitVal = dyn_cast_or_null<ConstantInt>(Val: IRInstrVar->getInitializer()); |
| 1511 | if (!InitVal) |
| 1512 | return false; |
| 1513 | return (InitVal->getZExtValue() & VARIANT_MASK_IR_PROF) != 0; |
| 1514 | } |
| 1515 | |
| 1516 | // Check if we can safely rename this Comdat function. |
| 1517 | bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) { |
| 1518 | if (F.getName().empty()) |
| 1519 | return false; |
| 1520 | if (!needsComdatForCounter(GO: F, M: *(F.getParent()))) |
| 1521 | return false; |
| 1522 | // Unsafe to rename the address-taken function (which can be used in |
| 1523 | // function comparison). |
| 1524 | if (CheckAddressTaken && F.hasAddressTaken()) |
| 1525 | return false; |
| 1526 | // Only safe to do if this function may be discarded if it is not used |
| 1527 | // in the compilation unit. |
| 1528 | if (!GlobalValue::isDiscardableIfUnused(Linkage: F.getLinkage())) |
| 1529 | return false; |
| 1530 | |
| 1531 | // For AvailableExternallyLinkage functions. |
| 1532 | if (!F.hasComdat()) { |
| 1533 | assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage); |
| 1534 | return true; |
| 1535 | } |
| 1536 | return true; |
| 1537 | } |
| 1538 | |
| 1539 | // Create the variable for the profile file name. |
| 1540 | void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput) { |
| 1541 | if (InstrProfileOutput.empty()) |
| 1542 | return; |
| 1543 | Constant *ProfileNameConst = |
| 1544 | ConstantDataArray::getString(Context&: M.getContext(), Initializer: InstrProfileOutput, AddNull: true); |
| 1545 | GlobalVariable *ProfileNameVar = new GlobalVariable( |
| 1546 | M, ProfileNameConst->getType(), true, GlobalValue::WeakAnyLinkage, |
| 1547 | ProfileNameConst, INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR)); |
| 1548 | ProfileNameVar->setVisibility(GlobalValue::HiddenVisibility); |
| 1549 | Triple TT(M.getTargetTriple()); |
| 1550 | if (TT.supportsCOMDAT()) { |
| 1551 | ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage); |
| 1552 | ProfileNameVar->setComdat(M.getOrInsertComdat( |
| 1553 | Name: StringRef(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR)))); |
| 1554 | } |
| 1555 | } |
| 1556 | |
| 1557 | Error OverlapStats::accumulateCounts(const std::string &BaseFilename, |
| 1558 | const std::string &TestFilename, |
| 1559 | bool IsCS) { |
| 1560 | auto GetProfileSum = [IsCS](const std::string &Filename, |
| 1561 | CountSumOrPercent &Sum) -> Error { |
| 1562 | // This function is only used from llvm-profdata that doesn't use any kind |
| 1563 | // of VFS. Just create a default RealFileSystem to read profiles. |
| 1564 | auto FS = vfs::getRealFileSystem(); |
| 1565 | auto ReaderOrErr = InstrProfReader::create(Path: Filename, FS&: *FS); |
| 1566 | if (Error E = ReaderOrErr.takeError()) { |
| 1567 | return E; |
| 1568 | } |
| 1569 | auto Reader = std::move(ReaderOrErr.get()); |
| 1570 | Reader->accumulateCounts(Sum, IsCS); |
| 1571 | return Error::success(); |
| 1572 | }; |
| 1573 | auto Ret = GetProfileSum(BaseFilename, Base); |
| 1574 | if (Ret) |
| 1575 | return Ret; |
| 1576 | Ret = GetProfileSum(TestFilename, Test); |
| 1577 | if (Ret) |
| 1578 | return Ret; |
| 1579 | this->BaseFilename = &BaseFilename; |
| 1580 | this->TestFilename = &TestFilename; |
| 1581 | Valid = true; |
| 1582 | return Error::success(); |
| 1583 | } |
| 1584 | |
| 1585 | void OverlapStats::addOneMismatch(const CountSumOrPercent &MismatchFunc) { |
| 1586 | Mismatch.NumEntries += 1; |
| 1587 | Mismatch.CountSum += MismatchFunc.CountSum / Test.CountSum; |
| 1588 | for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { |
| 1589 | if (Test.ValueCounts[I] >= 1.0f) |
| 1590 | Mismatch.ValueCounts[I] += |
| 1591 | MismatchFunc.ValueCounts[I] / Test.ValueCounts[I]; |
| 1592 | } |
| 1593 | } |
| 1594 | |
| 1595 | void OverlapStats::addOneUnique(const CountSumOrPercent &UniqueFunc) { |
| 1596 | Unique.NumEntries += 1; |
| 1597 | Unique.CountSum += UniqueFunc.CountSum / Test.CountSum; |
| 1598 | for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { |
| 1599 | if (Test.ValueCounts[I] >= 1.0f) |
| 1600 | Unique.ValueCounts[I] += UniqueFunc.ValueCounts[I] / Test.ValueCounts[I]; |
| 1601 | } |
| 1602 | } |
| 1603 | |
| 1604 | void OverlapStats::dump(raw_fd_ostream &OS) const { |
| 1605 | if (!Valid) |
| 1606 | return; |
| 1607 | |
| 1608 | const char *EntryName = |
| 1609 | (Level == ProgramLevel ? "functions" : "edge counters" ); |
| 1610 | if (Level == ProgramLevel) { |
| 1611 | OS << "Profile overlap infomation for base_profile: " << *BaseFilename |
| 1612 | << " and test_profile: " << *TestFilename << "\nProgram level:\n" ; |
| 1613 | } else { |
| 1614 | OS << "Function level:\n" |
| 1615 | << " Function: " << FuncName << " (Hash=" << FuncHash << ")\n" ; |
| 1616 | } |
| 1617 | |
| 1618 | OS << " # of " << EntryName << " overlap: " << Overlap.NumEntries << "\n" ; |
| 1619 | if (Mismatch.NumEntries) |
| 1620 | OS << " # of " << EntryName << " mismatch: " << Mismatch.NumEntries |
| 1621 | << "\n" ; |
| 1622 | if (Unique.NumEntries) |
| 1623 | OS << " # of " << EntryName |
| 1624 | << " only in test_profile: " << Unique.NumEntries << "\n" ; |
| 1625 | |
| 1626 | OS << " Edge profile overlap: " << format(Fmt: "%.3f%%" , Vals: Overlap.CountSum * 100) |
| 1627 | << "\n" ; |
| 1628 | if (Mismatch.NumEntries) |
| 1629 | OS << " Mismatched count percentage (Edge): " |
| 1630 | << format(Fmt: "%.3f%%" , Vals: Mismatch.CountSum * 100) << "\n" ; |
| 1631 | if (Unique.NumEntries) |
| 1632 | OS << " Percentage of Edge profile only in test_profile: " |
| 1633 | << format(Fmt: "%.3f%%" , Vals: Unique.CountSum * 100) << "\n" ; |
| 1634 | OS << " Edge profile base count sum: " << format(Fmt: "%.0f" , Vals: Base.CountSum) |
| 1635 | << "\n" |
| 1636 | << " Edge profile test count sum: " << format(Fmt: "%.0f" , Vals: Test.CountSum) |
| 1637 | << "\n" ; |
| 1638 | |
| 1639 | for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { |
| 1640 | if (Base.ValueCounts[I] < 1.0f && Test.ValueCounts[I] < 1.0f) |
| 1641 | continue; |
| 1642 | char ProfileKindName[20] = {0}; |
| 1643 | switch (I) { |
| 1644 | case IPVK_IndirectCallTarget: |
| 1645 | strncpy(dest: ProfileKindName, src: "IndirectCall" , n: 19); |
| 1646 | break; |
| 1647 | case IPVK_MemOPSize: |
| 1648 | strncpy(dest: ProfileKindName, src: "MemOP" , n: 19); |
| 1649 | break; |
| 1650 | case IPVK_VTableTarget: |
| 1651 | strncpy(dest: ProfileKindName, src: "VTable" , n: 19); |
| 1652 | break; |
| 1653 | default: |
| 1654 | snprintf(s: ProfileKindName, maxlen: 19, format: "VP[%d]" , I); |
| 1655 | break; |
| 1656 | } |
| 1657 | OS << " " << ProfileKindName |
| 1658 | << " profile overlap: " << format(Fmt: "%.3f%%" , Vals: Overlap.ValueCounts[I] * 100) |
| 1659 | << "\n" ; |
| 1660 | if (Mismatch.NumEntries) |
| 1661 | OS << " Mismatched count percentage (" << ProfileKindName |
| 1662 | << "): " << format(Fmt: "%.3f%%" , Vals: Mismatch.ValueCounts[I] * 100) << "\n" ; |
| 1663 | if (Unique.NumEntries) |
| 1664 | OS << " Percentage of " << ProfileKindName |
| 1665 | << " profile only in test_profile: " |
| 1666 | << format(Fmt: "%.3f%%" , Vals: Unique.ValueCounts[I] * 100) << "\n" ; |
| 1667 | OS << " " << ProfileKindName |
| 1668 | << " profile base count sum: " << format(Fmt: "%.0f" , Vals: Base.ValueCounts[I]) |
| 1669 | << "\n" |
| 1670 | << " " << ProfileKindName |
| 1671 | << " profile test count sum: " << format(Fmt: "%.0f" , Vals: Test.ValueCounts[I]) |
| 1672 | << "\n" ; |
| 1673 | } |
| 1674 | } |
| 1675 | |
| 1676 | namespace IndexedInstrProf { |
| 1677 | Expected<Header> Header::(const unsigned char *Buffer) { |
| 1678 | using namespace support; |
| 1679 | static_assert(std::is_standard_layout_v<Header>, |
| 1680 | "Use standard layout for Header for simplicity" ); |
| 1681 | Header H; |
| 1682 | |
| 1683 | H.Magic = endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1684 | // Check the magic number. |
| 1685 | if (H.Magic != IndexedInstrProf::Magic) |
| 1686 | return make_error<InstrProfError>(Args: instrprof_error::bad_magic); |
| 1687 | |
| 1688 | // Read the version. |
| 1689 | H.Version = endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1690 | if (H.getIndexedProfileVersion() > |
| 1691 | IndexedInstrProf::ProfVersion::CurrentVersion) |
| 1692 | return make_error<InstrProfError>(Args: instrprof_error::unsupported_version); |
| 1693 | |
| 1694 | static_assert(IndexedInstrProf::ProfVersion::CurrentVersion == Version12, |
| 1695 | "Please update the reader as needed when a new field is added " |
| 1696 | "or when indexed profile version gets bumped." ); |
| 1697 | |
| 1698 | Buffer += sizeof(uint64_t); // Skip Header.Unused field. |
| 1699 | H.HashType = endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1700 | H.HashOffset = endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1701 | if (H.getIndexedProfileVersion() >= 8) |
| 1702 | H.MemProfOffset = |
| 1703 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1704 | if (H.getIndexedProfileVersion() >= 9) |
| 1705 | H.BinaryIdOffset = |
| 1706 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1707 | // Version 11 is handled by this condition. |
| 1708 | if (H.getIndexedProfileVersion() >= 10) |
| 1709 | H.TemporalProfTracesOffset = |
| 1710 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1711 | if (H.getIndexedProfileVersion() >= 12) |
| 1712 | H.VTableNamesOffset = |
| 1713 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: Buffer); |
| 1714 | return H; |
| 1715 | } |
| 1716 | |
| 1717 | uint64_t Header::() const { |
| 1718 | return GET_VERSION(Version); |
| 1719 | } |
| 1720 | |
| 1721 | size_t Header::() const { |
| 1722 | switch (getIndexedProfileVersion()) { |
| 1723 | // To retain backward compatibility, new fields must be appended to the end |
| 1724 | // of the header, and byte offset of existing fields shouldn't change when |
| 1725 | // indexed profile version gets incremented. |
| 1726 | static_assert( |
| 1727 | IndexedInstrProf::ProfVersion::CurrentVersion == Version12, |
| 1728 | "Please update the size computation below if a new field has " |
| 1729 | "been added to the header; for a version bump without new " |
| 1730 | "fields, add a case statement to fall through to the latest version." ); |
| 1731 | case 12ull: |
| 1732 | return 72; |
| 1733 | case 11ull: |
| 1734 | [[fallthrough]]; |
| 1735 | case 10ull: |
| 1736 | return 64; |
| 1737 | case 9ull: |
| 1738 | return 56; |
| 1739 | case 8ull: |
| 1740 | return 48; |
| 1741 | default: // Version7 (when the backwards compatible header was introduced). |
| 1742 | return 40; |
| 1743 | } |
| 1744 | } |
| 1745 | |
| 1746 | } // namespace IndexedInstrProf |
| 1747 | |
| 1748 | } // end namespace llvm |
| 1749 | |