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/SmallString.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/StringExtras.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/ADT/Triple.h" 21 #include "llvm/IR/Constant.h" 22 #include "llvm/IR/Constants.h" 23 #include "llvm/IR/Function.h" 24 #include "llvm/IR/GlobalValue.h" 25 #include "llvm/IR/GlobalVariable.h" 26 #include "llvm/IR/Instruction.h" 27 #include "llvm/IR/LLVMContext.h" 28 #include "llvm/IR/MDBuilder.h" 29 #include "llvm/IR/Metadata.h" 30 #include "llvm/IR/Module.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/Endian.h" 38 #include "llvm/Support/Error.h" 39 #include "llvm/Support/ErrorHandling.h" 40 #include "llvm/Support/LEB128.h" 41 #include "llvm/Support/ManagedStatic.h" 42 #include "llvm/Support/MathExtras.h" 43 #include "llvm/Support/Path.h" 44 #include "llvm/Support/SwapByteOrder.h" 45 #include <algorithm> 46 #include <cassert> 47 #include <cstddef> 48 #include <cstdint> 49 #include <cstring> 50 #include <memory> 51 #include <string> 52 #include <system_error> 53 #include <utility> 54 #include <vector> 55 56 using namespace llvm; 57 58 static cl::opt<bool> StaticFuncFullModulePrefix( 59 "static-func-full-module-prefix", cl::init(true), cl::Hidden, 60 cl::desc("Use full module build paths in the profile counter names for " 61 "static functions.")); 62 63 // This option is tailored to users that have different top-level directory in 64 // profile-gen and profile-use compilation. Users need to specific the number 65 // of levels to strip. A value larger than the number of directories in the 66 // source file will strip all the directory names and only leave the basename. 67 // 68 // Note current ThinLTO module importing for the indirect-calls assumes 69 // the source directory name not being stripped. A non-zero option value here 70 // can potentially prevent some inter-module indirect-call-promotions. 71 static cl::opt<unsigned> StaticFuncStripDirNamePrefix( 72 "static-func-strip-dirname-prefix", cl::init(0), cl::Hidden, 73 cl::desc("Strip specified level of directory name from source path in " 74 "the profile counter name for static functions.")); 75 76 static std::string getInstrProfErrString(instrprof_error Err) { 77 switch (Err) { 78 case instrprof_error::success: 79 return "Success"; 80 case instrprof_error::eof: 81 return "End of File"; 82 case instrprof_error::unrecognized_format: 83 return "Unrecognized instrumentation profile encoding format"; 84 case instrprof_error::bad_magic: 85 return "Invalid instrumentation profile data (bad magic)"; 86 case instrprof_error::bad_header: 87 return "Invalid instrumentation profile data (file header is corrupt)"; 88 case instrprof_error::unsupported_version: 89 return "Unsupported instrumentation profile format version"; 90 case instrprof_error::unsupported_hash_type: 91 return "Unsupported instrumentation profile hash type"; 92 case instrprof_error::too_large: 93 return "Too much profile data"; 94 case instrprof_error::truncated: 95 return "Truncated profile data"; 96 case instrprof_error::malformed: 97 return "Malformed instrumentation profile data"; 98 case instrprof_error::unknown_function: 99 return "No profile data available for function"; 100 case instrprof_error::hash_mismatch: 101 return "Function control flow change detected (hash mismatch)"; 102 case instrprof_error::count_mismatch: 103 return "Function basic block count change detected (counter mismatch)"; 104 case instrprof_error::counter_overflow: 105 return "Counter overflow"; 106 case instrprof_error::value_site_count_mismatch: 107 return "Function value site count change detected (counter mismatch)"; 108 case instrprof_error::compress_failed: 109 return "Failed to compress data (zlib)"; 110 case instrprof_error::uncompress_failed: 111 return "Failed to uncompress data (zlib)"; 112 case instrprof_error::empty_raw_profile: 113 return "Empty raw profile file"; 114 case instrprof_error::zlib_unavailable: 115 return "Profile uses zlib compression but the profile reader was built without zlib support"; 116 } 117 llvm_unreachable("A value of instrprof_error has no message."); 118 } 119 120 namespace { 121 122 // FIXME: This class is only here to support the transition to llvm::Error. It 123 // will be removed once this transition is complete. Clients should prefer to 124 // deal with the Error value directly, rather than converting to error_code. 125 class InstrProfErrorCategoryType : public std::error_category { 126 const char *name() const noexcept override { return "llvm.instrprof"; } 127 128 std::string message(int IE) const override { 129 return getInstrProfErrString(static_cast<instrprof_error>(IE)); 130 } 131 }; 132 133 } // end anonymous namespace 134 135 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory; 136 137 const std::error_category &llvm::instrprof_category() { 138 return *ErrorCategory; 139 } 140 141 namespace { 142 143 const char *InstrProfSectNameCommon[] = { 144 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ 145 SectNameCommon, 146 #include "llvm/ProfileData/InstrProfData.inc" 147 }; 148 149 const char *InstrProfSectNameCoff[] = { 150 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ 151 SectNameCoff, 152 #include "llvm/ProfileData/InstrProfData.inc" 153 }; 154 155 const char *InstrProfSectNamePrefix[] = { 156 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) \ 157 Prefix, 158 #include "llvm/ProfileData/InstrProfData.inc" 159 }; 160 161 } // namespace 162 163 namespace llvm { 164 165 std::string getInstrProfSectionName(InstrProfSectKind IPSK, 166 Triple::ObjectFormatType OF, 167 bool AddSegmentInfo) { 168 std::string SectName; 169 170 if (OF == Triple::MachO && AddSegmentInfo) 171 SectName = InstrProfSectNamePrefix[IPSK]; 172 173 if (OF == Triple::COFF) 174 SectName += InstrProfSectNameCoff[IPSK]; 175 else 176 SectName += InstrProfSectNameCommon[IPSK]; 177 178 if (OF == Triple::MachO && IPSK == IPSK_data && AddSegmentInfo) 179 SectName += ",regular,live_support"; 180 181 return SectName; 182 } 183 184 void SoftInstrProfErrors::addError(instrprof_error IE) { 185 if (IE == instrprof_error::success) 186 return; 187 188 if (FirstError == instrprof_error::success) 189 FirstError = IE; 190 191 switch (IE) { 192 case instrprof_error::hash_mismatch: 193 ++NumHashMismatches; 194 break; 195 case instrprof_error::count_mismatch: 196 ++NumCountMismatches; 197 break; 198 case instrprof_error::counter_overflow: 199 ++NumCounterOverflows; 200 break; 201 case instrprof_error::value_site_count_mismatch: 202 ++NumValueSiteCountMismatches; 203 break; 204 default: 205 llvm_unreachable("Not a soft error"); 206 } 207 } 208 209 std::string InstrProfError::message() const { 210 return getInstrProfErrString(Err); 211 } 212 213 char InstrProfError::ID = 0; 214 215 std::string getPGOFuncName(StringRef RawFuncName, 216 GlobalValue::LinkageTypes Linkage, 217 StringRef FileName, 218 uint64_t Version LLVM_ATTRIBUTE_UNUSED) { 219 return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName); 220 } 221 222 // Strip NumPrefix level of directory name from PathNameStr. If the number of 223 // directory separators is less than NumPrefix, strip all the directories and 224 // leave base file name only. 225 static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) { 226 uint32_t Count = NumPrefix; 227 uint32_t Pos = 0, LastPos = 0; 228 for (auto & CI : PathNameStr) { 229 ++Pos; 230 if (llvm::sys::path::is_separator(CI)) { 231 LastPos = Pos; 232 --Count; 233 } 234 if (Count == 0) 235 break; 236 } 237 return PathNameStr.substr(LastPos); 238 } 239 240 // Return the PGOFuncName. This function has some special handling when called 241 // in LTO optimization. The following only applies when calling in LTO passes 242 // (when \c InLTO is true): LTO's internalization privatizes many global linkage 243 // symbols. This happens after value profile annotation, but those internal 244 // linkage functions should not have a source prefix. 245 // Additionally, for ThinLTO mode, exported internal functions are promoted 246 // and renamed. We need to ensure that the original internal PGO name is 247 // used when computing the GUID that is compared against the profiled GUIDs. 248 // To differentiate compiler generated internal symbols from original ones, 249 // PGOFuncName meta data are created and attached to the original internal 250 // symbols in the value profile annotation step 251 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta 252 // data, its original linkage must be non-internal. 253 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) { 254 if (!InLTO) { 255 StringRef FileName(F.getParent()->getSourceFileName()); 256 uint32_t StripLevel = StaticFuncFullModulePrefix ? 0 : (uint32_t)-1; 257 if (StripLevel < StaticFuncStripDirNamePrefix) 258 StripLevel = StaticFuncStripDirNamePrefix; 259 if (StripLevel) 260 FileName = stripDirPrefix(FileName, StripLevel); 261 return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version); 262 } 263 264 // In LTO mode (when InLTO is true), first check if there is a meta data. 265 if (MDNode *MD = getPGOFuncNameMetadata(F)) { 266 StringRef S = cast<MDString>(MD->getOperand(0))->getString(); 267 return S.str(); 268 } 269 270 // If there is no meta data, the function must be a global before the value 271 // profile annotation pass. Its current linkage may be internal if it is 272 // internalized in LTO mode. 273 return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, ""); 274 } 275 276 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) { 277 if (FileName.empty()) 278 return PGOFuncName; 279 // Drop the file name including ':'. See also getPGOFuncName. 280 if (PGOFuncName.startswith(FileName)) 281 PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1); 282 return PGOFuncName; 283 } 284 285 // \p FuncName is the string used as profile lookup key for the function. A 286 // symbol is created to hold the name. Return the legalized symbol name. 287 std::string getPGOFuncNameVarName(StringRef FuncName, 288 GlobalValue::LinkageTypes Linkage) { 289 std::string VarName = getInstrProfNameVarPrefix(); 290 VarName += FuncName; 291 292 if (!GlobalValue::isLocalLinkage(Linkage)) 293 return VarName; 294 295 // Now fix up illegal chars in local VarName that may upset the assembler. 296 const char *InvalidChars = "-:<>/\"'"; 297 size_t found = VarName.find_first_of(InvalidChars); 298 while (found != std::string::npos) { 299 VarName[found] = '_'; 300 found = VarName.find_first_of(InvalidChars, found + 1); 301 } 302 return VarName; 303 } 304 305 GlobalVariable *createPGOFuncNameVar(Module &M, 306 GlobalValue::LinkageTypes Linkage, 307 StringRef PGOFuncName) { 308 // We generally want to match the function's linkage, but available_externally 309 // and extern_weak both have the wrong semantics, and anything that doesn't 310 // need to link across compilation units doesn't need to be visible at all. 311 if (Linkage == GlobalValue::ExternalWeakLinkage) 312 Linkage = GlobalValue::LinkOnceAnyLinkage; 313 else if (Linkage == GlobalValue::AvailableExternallyLinkage) 314 Linkage = GlobalValue::LinkOnceODRLinkage; 315 else if (Linkage == GlobalValue::InternalLinkage || 316 Linkage == GlobalValue::ExternalLinkage) 317 Linkage = GlobalValue::PrivateLinkage; 318 319 auto *Value = 320 ConstantDataArray::getString(M.getContext(), PGOFuncName, false); 321 auto FuncNameVar = 322 new GlobalVariable(M, Value->getType(), true, Linkage, Value, 323 getPGOFuncNameVarName(PGOFuncName, Linkage)); 324 325 // Hide the symbol so that we correctly get a copy for each executable. 326 if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage())) 327 FuncNameVar->setVisibility(GlobalValue::HiddenVisibility); 328 329 return FuncNameVar; 330 } 331 332 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) { 333 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName); 334 } 335 336 Error InstrProfSymtab::create(Module &M, bool InLTO) { 337 for (Function &F : M) { 338 // Function may not have a name: like using asm("") to overwrite the name. 339 // Ignore in this case. 340 if (!F.hasName()) 341 continue; 342 const std::string &PGOFuncName = getPGOFuncName(F, InLTO); 343 if (Error E = addFuncName(PGOFuncName)) 344 return E; 345 MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F); 346 // In ThinLTO, local function may have been promoted to global and have 347 // suffix added to the function name. We need to add the stripped function 348 // name to the symbol table so that we can find a match from profile. 349 if (InLTO) { 350 auto pos = PGOFuncName.find('.'); 351 if (pos != std::string::npos) { 352 const std::string &OtherFuncName = PGOFuncName.substr(0, pos); 353 if (Error E = addFuncName(OtherFuncName)) 354 return E; 355 MD5FuncMap.emplace_back(Function::getGUID(OtherFuncName), &F); 356 } 357 } 358 } 359 Sorted = false; 360 finalizeSymtab(); 361 return Error::success(); 362 } 363 364 uint64_t InstrProfSymtab::getFunctionHashFromAddress(uint64_t Address) { 365 finalizeSymtab(); 366 auto Result = 367 std::lower_bound(AddrToMD5Map.begin(), AddrToMD5Map.end(), Address, 368 [](const std::pair<uint64_t, uint64_t> &LHS, 369 uint64_t RHS) { return LHS.first < RHS; }); 370 // Raw function pointer collected by value profiler may be from 371 // external functions that are not instrumented. They won't have 372 // mapping data to be used by the deserializer. Force the value to 373 // be 0 in this case. 374 if (Result != AddrToMD5Map.end() && Result->first == Address) 375 return (uint64_t)Result->second; 376 return 0; 377 } 378 379 Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs, 380 bool doCompression, std::string &Result) { 381 assert(!NameStrs.empty() && "No name data to emit"); 382 383 uint8_t Header[16], *P = Header; 384 std::string UncompressedNameStrings = 385 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator()); 386 387 assert(StringRef(UncompressedNameStrings) 388 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) && 389 "PGO name is invalid (contains separator token)"); 390 391 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P); 392 P += EncLen; 393 394 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) { 395 EncLen = encodeULEB128(CompressedLen, P); 396 P += EncLen; 397 char *HeaderStr = reinterpret_cast<char *>(&Header[0]); 398 unsigned HeaderLen = P - &Header[0]; 399 Result.append(HeaderStr, HeaderLen); 400 Result += InputStr; 401 return Error::success(); 402 }; 403 404 if (!doCompression) { 405 return WriteStringToResult(0, UncompressedNameStrings); 406 } 407 408 SmallString<128> CompressedNameStrings; 409 Error E = zlib::compress(StringRef(UncompressedNameStrings), 410 CompressedNameStrings, zlib::BestSizeCompression); 411 if (E) { 412 consumeError(std::move(E)); 413 return make_error<InstrProfError>(instrprof_error::compress_failed); 414 } 415 416 return WriteStringToResult(CompressedNameStrings.size(), 417 CompressedNameStrings); 418 } 419 420 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) { 421 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer()); 422 StringRef NameStr = 423 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString(); 424 return NameStr; 425 } 426 427 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars, 428 std::string &Result, bool doCompression) { 429 std::vector<std::string> NameStrs; 430 for (auto *NameVar : NameVars) { 431 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar)); 432 } 433 return collectPGOFuncNameStrings( 434 NameStrs, zlib::isAvailable() && doCompression, Result); 435 } 436 437 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) { 438 const uint8_t *P = NameStrings.bytes_begin(); 439 const uint8_t *EndP = NameStrings.bytes_end(); 440 while (P < EndP) { 441 uint32_t N; 442 uint64_t UncompressedSize = decodeULEB128(P, &N); 443 P += N; 444 uint64_t CompressedSize = decodeULEB128(P, &N); 445 P += N; 446 bool isCompressed = (CompressedSize != 0); 447 SmallString<128> UncompressedNameStrings; 448 StringRef NameStrings; 449 if (isCompressed) { 450 if (!llvm::zlib::isAvailable()) 451 return make_error<InstrProfError>(instrprof_error::zlib_unavailable); 452 453 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P), 454 CompressedSize); 455 if (Error E = 456 zlib::uncompress(CompressedNameStrings, UncompressedNameStrings, 457 UncompressedSize)) { 458 consumeError(std::move(E)); 459 return make_error<InstrProfError>(instrprof_error::uncompress_failed); 460 } 461 P += CompressedSize; 462 NameStrings = StringRef(UncompressedNameStrings.data(), 463 UncompressedNameStrings.size()); 464 } else { 465 NameStrings = 466 StringRef(reinterpret_cast<const char *>(P), UncompressedSize); 467 P += UncompressedSize; 468 } 469 // Now parse the name strings. 470 SmallVector<StringRef, 0> Names; 471 NameStrings.split(Names, getInstrProfNameSeparator()); 472 for (StringRef &Name : Names) 473 if (Error E = Symtab.addFuncName(Name)) 474 return E; 475 476 while (P < EndP && *P == 0) 477 P++; 478 } 479 return Error::success(); 480 } 481 482 void InstrProfRecord::accumuateCounts(CountSumOrPercent &Sum) const { 483 uint64_t FuncSum = 0; 484 Sum.NumEntries += Counts.size(); 485 for (size_t F = 0, E = Counts.size(); F < E; ++F) 486 FuncSum += Counts[F]; 487 Sum.CountSum += FuncSum; 488 489 for (uint32_t VK = IPVK_First; VK <= IPVK_Last; ++VK) { 490 uint64_t KindSum = 0; 491 uint32_t NumValueSites = getNumValueSites(VK); 492 for (size_t I = 0; I < NumValueSites; ++I) { 493 uint32_t NV = getNumValueDataForSite(VK, I); 494 std::unique_ptr<InstrProfValueData[]> VD = getValueForSite(VK, I); 495 for (uint32_t V = 0; V < NV; V++) 496 KindSum += VD[V].Count; 497 } 498 Sum.ValueCounts[VK] += KindSum; 499 } 500 } 501 502 void InstrProfValueSiteRecord::overlap(InstrProfValueSiteRecord &Input, 503 uint32_t ValueKind, 504 OverlapStats &Overlap, 505 OverlapStats &FuncLevelOverlap) { 506 this->sortByTargetValues(); 507 Input.sortByTargetValues(); 508 double Score = 0.0f, FuncLevelScore = 0.0f; 509 auto I = ValueData.begin(); 510 auto IE = ValueData.end(); 511 auto J = Input.ValueData.begin(); 512 auto JE = Input.ValueData.end(); 513 while (I != IE && J != JE) { 514 if (I->Value == J->Value) { 515 Score += OverlapStats::score(I->Count, J->Count, 516 Overlap.Base.ValueCounts[ValueKind], 517 Overlap.Test.ValueCounts[ValueKind]); 518 FuncLevelScore += OverlapStats::score( 519 I->Count, J->Count, FuncLevelOverlap.Base.ValueCounts[ValueKind], 520 FuncLevelOverlap.Test.ValueCounts[ValueKind]); 521 ++I; 522 } else if (I->Value < J->Value) { 523 ++I; 524 continue; 525 } 526 ++J; 527 } 528 Overlap.Overlap.ValueCounts[ValueKind] += Score; 529 FuncLevelOverlap.Overlap.ValueCounts[ValueKind] += FuncLevelScore; 530 } 531 532 // Return false on mismatch. 533 void InstrProfRecord::overlapValueProfData(uint32_t ValueKind, 534 InstrProfRecord &Other, 535 OverlapStats &Overlap, 536 OverlapStats &FuncLevelOverlap) { 537 uint32_t ThisNumValueSites = getNumValueSites(ValueKind); 538 assert(ThisNumValueSites == Other.getNumValueSites(ValueKind)); 539 if (!ThisNumValueSites) 540 return; 541 542 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords = 543 getOrCreateValueSitesForKind(ValueKind); 544 MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords = 545 Other.getValueSitesForKind(ValueKind); 546 for (uint32_t I = 0; I < ThisNumValueSites; I++) 547 ThisSiteRecords[I].overlap(OtherSiteRecords[I], ValueKind, Overlap, 548 FuncLevelOverlap); 549 } 550 551 void InstrProfRecord::overlap(InstrProfRecord &Other, OverlapStats &Overlap, 552 OverlapStats &FuncLevelOverlap, 553 uint64_t ValueCutoff) { 554 // FuncLevel CountSum for other should already computed and nonzero. 555 assert(FuncLevelOverlap.Test.CountSum >= 1.0f); 556 accumuateCounts(FuncLevelOverlap.Base); 557 bool Mismatch = (Counts.size() != Other.Counts.size()); 558 559 // Check if the value profiles mismatch. 560 if (!Mismatch) { 561 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) { 562 uint32_t ThisNumValueSites = getNumValueSites(Kind); 563 uint32_t OtherNumValueSites = Other.getNumValueSites(Kind); 564 if (ThisNumValueSites != OtherNumValueSites) { 565 Mismatch = true; 566 break; 567 } 568 } 569 } 570 if (Mismatch) { 571 Overlap.addOneMismatch(FuncLevelOverlap.Test); 572 return; 573 } 574 575 // Compute overlap for value counts. 576 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) 577 overlapValueProfData(Kind, Other, Overlap, FuncLevelOverlap); 578 579 double Score = 0.0; 580 uint64_t MaxCount = 0; 581 // Compute overlap for edge counts. 582 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) { 583 Score += OverlapStats::score(Counts[I], Other.Counts[I], 584 Overlap.Base.CountSum, Overlap.Test.CountSum); 585 MaxCount = std::max(Other.Counts[I], MaxCount); 586 } 587 Overlap.Overlap.CountSum += Score; 588 Overlap.Overlap.NumEntries += 1; 589 590 if (MaxCount >= ValueCutoff) { 591 double FuncScore = 0.0; 592 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) 593 FuncScore += OverlapStats::score(Counts[I], Other.Counts[I], 594 FuncLevelOverlap.Base.CountSum, 595 FuncLevelOverlap.Test.CountSum); 596 FuncLevelOverlap.Overlap.CountSum = FuncScore; 597 FuncLevelOverlap.Overlap.NumEntries = Other.Counts.size(); 598 FuncLevelOverlap.Valid = true; 599 } 600 } 601 602 void InstrProfValueSiteRecord::merge(InstrProfValueSiteRecord &Input, 603 uint64_t Weight, 604 function_ref<void(instrprof_error)> Warn) { 605 this->sortByTargetValues(); 606 Input.sortByTargetValues(); 607 auto I = ValueData.begin(); 608 auto IE = ValueData.end(); 609 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE; 610 ++J) { 611 while (I != IE && I->Value < J->Value) 612 ++I; 613 if (I != IE && I->Value == J->Value) { 614 bool Overflowed; 615 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed); 616 if (Overflowed) 617 Warn(instrprof_error::counter_overflow); 618 ++I; 619 continue; 620 } 621 ValueData.insert(I, *J); 622 } 623 } 624 625 void InstrProfValueSiteRecord::scale(uint64_t Weight, 626 function_ref<void(instrprof_error)> Warn) { 627 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) { 628 bool Overflowed; 629 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed); 630 if (Overflowed) 631 Warn(instrprof_error::counter_overflow); 632 } 633 } 634 635 // Merge Value Profile data from Src record to this record for ValueKind. 636 // Scale merged value counts by \p Weight. 637 void InstrProfRecord::mergeValueProfData( 638 uint32_t ValueKind, InstrProfRecord &Src, uint64_t Weight, 639 function_ref<void(instrprof_error)> Warn) { 640 uint32_t ThisNumValueSites = getNumValueSites(ValueKind); 641 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind); 642 if (ThisNumValueSites != OtherNumValueSites) { 643 Warn(instrprof_error::value_site_count_mismatch); 644 return; 645 } 646 if (!ThisNumValueSites) 647 return; 648 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords = 649 getOrCreateValueSitesForKind(ValueKind); 650 MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords = 651 Src.getValueSitesForKind(ValueKind); 652 for (uint32_t I = 0; I < ThisNumValueSites; I++) 653 ThisSiteRecords[I].merge(OtherSiteRecords[I], Weight, Warn); 654 } 655 656 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight, 657 function_ref<void(instrprof_error)> Warn) { 658 // If the number of counters doesn't match we either have bad data 659 // or a hash collision. 660 if (Counts.size() != Other.Counts.size()) { 661 Warn(instrprof_error::count_mismatch); 662 return; 663 } 664 665 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) { 666 bool Overflowed; 667 Counts[I] = 668 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed); 669 if (Overflowed) 670 Warn(instrprof_error::counter_overflow); 671 } 672 673 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) 674 mergeValueProfData(Kind, Other, Weight, Warn); 675 } 676 677 void InstrProfRecord::scaleValueProfData( 678 uint32_t ValueKind, uint64_t Weight, 679 function_ref<void(instrprof_error)> Warn) { 680 for (auto &R : getValueSitesForKind(ValueKind)) 681 R.scale(Weight, Warn); 682 } 683 684 void InstrProfRecord::scale(uint64_t Weight, 685 function_ref<void(instrprof_error)> Warn) { 686 for (auto &Count : this->Counts) { 687 bool Overflowed; 688 Count = SaturatingMultiply(Count, Weight, &Overflowed); 689 if (Overflowed) 690 Warn(instrprof_error::counter_overflow); 691 } 692 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) 693 scaleValueProfData(Kind, Weight, Warn); 694 } 695 696 // Map indirect call target name hash to name string. 697 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind, 698 InstrProfSymtab *SymTab) { 699 if (!SymTab) 700 return Value; 701 702 if (ValueKind == IPVK_IndirectCallTarget) 703 return SymTab->getFunctionHashFromAddress(Value); 704 705 return Value; 706 } 707 708 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site, 709 InstrProfValueData *VData, uint32_t N, 710 InstrProfSymtab *ValueMap) { 711 for (uint32_t I = 0; I < N; I++) { 712 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap); 713 } 714 std::vector<InstrProfValueSiteRecord> &ValueSites = 715 getOrCreateValueSitesForKind(ValueKind); 716 if (N == 0) 717 ValueSites.emplace_back(); 718 else 719 ValueSites.emplace_back(VData, VData + N); 720 } 721 722 #define INSTR_PROF_COMMON_API_IMPL 723 #include "llvm/ProfileData/InstrProfData.inc" 724 725 /*! 726 * ValueProfRecordClosure Interface implementation for InstrProfRecord 727 * class. These C wrappers are used as adaptors so that C++ code can be 728 * invoked as callbacks. 729 */ 730 uint32_t getNumValueKindsInstrProf(const void *Record) { 731 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds(); 732 } 733 734 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) { 735 return reinterpret_cast<const InstrProfRecord *>(Record) 736 ->getNumValueSites(VKind); 737 } 738 739 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) { 740 return reinterpret_cast<const InstrProfRecord *>(Record) 741 ->getNumValueData(VKind); 742 } 743 744 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK, 745 uint32_t S) { 746 return reinterpret_cast<const InstrProfRecord *>(R) 747 ->getNumValueDataForSite(VK, S); 748 } 749 750 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst, 751 uint32_t K, uint32_t S) { 752 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S); 753 } 754 755 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) { 756 ValueProfData *VD = 757 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData()); 758 memset(VD, 0, TotalSizeInBytes); 759 return VD; 760 } 761 762 static ValueProfRecordClosure InstrProfRecordClosure = { 763 nullptr, 764 getNumValueKindsInstrProf, 765 getNumValueSitesInstrProf, 766 getNumValueDataInstrProf, 767 getNumValueDataForSiteInstrProf, 768 nullptr, 769 getValueForSiteInstrProf, 770 allocValueProfDataInstrProf}; 771 772 // Wrapper implementation using the closure mechanism. 773 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) { 774 auto Closure = InstrProfRecordClosure; 775 Closure.Record = &Record; 776 return getValueProfDataSize(&Closure); 777 } 778 779 // Wrapper implementation using the closure mechanism. 780 std::unique_ptr<ValueProfData> 781 ValueProfData::serializeFrom(const InstrProfRecord &Record) { 782 InstrProfRecordClosure.Record = &Record; 783 784 std::unique_ptr<ValueProfData> VPD( 785 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr)); 786 return VPD; 787 } 788 789 void ValueProfRecord::deserializeTo(InstrProfRecord &Record, 790 InstrProfSymtab *SymTab) { 791 Record.reserveSites(Kind, NumValueSites); 792 793 InstrProfValueData *ValueData = getValueProfRecordValueData(this); 794 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) { 795 uint8_t ValueDataCount = this->SiteCountArray[VSite]; 796 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, SymTab); 797 ValueData += ValueDataCount; 798 } 799 } 800 801 // For writing/serializing, Old is the host endianness, and New is 802 // byte order intended on disk. For Reading/deserialization, Old 803 // is the on-disk source endianness, and New is the host endianness. 804 void ValueProfRecord::swapBytes(support::endianness Old, 805 support::endianness New) { 806 using namespace support; 807 808 if (Old == New) 809 return; 810 811 if (getHostEndianness() != Old) { 812 sys::swapByteOrder<uint32_t>(NumValueSites); 813 sys::swapByteOrder<uint32_t>(Kind); 814 } 815 uint32_t ND = getValueProfRecordNumValueData(this); 816 InstrProfValueData *VD = getValueProfRecordValueData(this); 817 818 // No need to swap byte array: SiteCountArrray. 819 for (uint32_t I = 0; I < ND; I++) { 820 sys::swapByteOrder<uint64_t>(VD[I].Value); 821 sys::swapByteOrder<uint64_t>(VD[I].Count); 822 } 823 if (getHostEndianness() == Old) { 824 sys::swapByteOrder<uint32_t>(NumValueSites); 825 sys::swapByteOrder<uint32_t>(Kind); 826 } 827 } 828 829 void ValueProfData::deserializeTo(InstrProfRecord &Record, 830 InstrProfSymtab *SymTab) { 831 if (NumValueKinds == 0) 832 return; 833 834 ValueProfRecord *VR = getFirstValueProfRecord(this); 835 for (uint32_t K = 0; K < NumValueKinds; K++) { 836 VR->deserializeTo(Record, SymTab); 837 VR = getValueProfRecordNext(VR); 838 } 839 } 840 841 template <class T> 842 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) { 843 using namespace support; 844 845 if (Orig == little) 846 return endian::readNext<T, little, unaligned>(D); 847 else 848 return endian::readNext<T, big, unaligned>(D); 849 } 850 851 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) { 852 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize)) 853 ValueProfData()); 854 } 855 856 Error ValueProfData::checkIntegrity() { 857 if (NumValueKinds > IPVK_Last + 1) 858 return make_error<InstrProfError>(instrprof_error::malformed); 859 // Total size needs to be mulltiple of quadword size. 860 if (TotalSize % sizeof(uint64_t)) 861 return make_error<InstrProfError>(instrprof_error::malformed); 862 863 ValueProfRecord *VR = getFirstValueProfRecord(this); 864 for (uint32_t K = 0; K < this->NumValueKinds; K++) { 865 if (VR->Kind > IPVK_Last) 866 return make_error<InstrProfError>(instrprof_error::malformed); 867 VR = getValueProfRecordNext(VR); 868 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize) 869 return make_error<InstrProfError>(instrprof_error::malformed); 870 } 871 return Error::success(); 872 } 873 874 Expected<std::unique_ptr<ValueProfData>> 875 ValueProfData::getValueProfData(const unsigned char *D, 876 const unsigned char *const BufferEnd, 877 support::endianness Endianness) { 878 using namespace support; 879 880 if (D + sizeof(ValueProfData) > BufferEnd) 881 return make_error<InstrProfError>(instrprof_error::truncated); 882 883 const unsigned char *Header = D; 884 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness); 885 if (D + TotalSize > BufferEnd) 886 return make_error<InstrProfError>(instrprof_error::too_large); 887 888 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize); 889 memcpy(VPD.get(), D, TotalSize); 890 // Byte swap. 891 VPD->swapBytesToHost(Endianness); 892 893 Error E = VPD->checkIntegrity(); 894 if (E) 895 return std::move(E); 896 897 return std::move(VPD); 898 } 899 900 void ValueProfData::swapBytesToHost(support::endianness Endianness) { 901 using namespace support; 902 903 if (Endianness == getHostEndianness()) 904 return; 905 906 sys::swapByteOrder<uint32_t>(TotalSize); 907 sys::swapByteOrder<uint32_t>(NumValueKinds); 908 909 ValueProfRecord *VR = getFirstValueProfRecord(this); 910 for (uint32_t K = 0; K < NumValueKinds; K++) { 911 VR->swapBytes(Endianness, getHostEndianness()); 912 VR = getValueProfRecordNext(VR); 913 } 914 } 915 916 void ValueProfData::swapBytesFromHost(support::endianness Endianness) { 917 using namespace support; 918 919 if (Endianness == getHostEndianness()) 920 return; 921 922 ValueProfRecord *VR = getFirstValueProfRecord(this); 923 for (uint32_t K = 0; K < NumValueKinds; K++) { 924 ValueProfRecord *NVR = getValueProfRecordNext(VR); 925 VR->swapBytes(getHostEndianness(), Endianness); 926 VR = NVR; 927 } 928 sys::swapByteOrder<uint32_t>(TotalSize); 929 sys::swapByteOrder<uint32_t>(NumValueKinds); 930 } 931 932 void annotateValueSite(Module &M, Instruction &Inst, 933 const InstrProfRecord &InstrProfR, 934 InstrProfValueKind ValueKind, uint32_t SiteIdx, 935 uint32_t MaxMDCount) { 936 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx); 937 if (!NV) 938 return; 939 940 uint64_t Sum = 0; 941 std::unique_ptr<InstrProfValueData[]> VD = 942 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum); 943 944 ArrayRef<InstrProfValueData> VDs(VD.get(), NV); 945 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount); 946 } 947 948 void annotateValueSite(Module &M, Instruction &Inst, 949 ArrayRef<InstrProfValueData> VDs, 950 uint64_t Sum, InstrProfValueKind ValueKind, 951 uint32_t MaxMDCount) { 952 LLVMContext &Ctx = M.getContext(); 953 MDBuilder MDHelper(Ctx); 954 SmallVector<Metadata *, 3> Vals; 955 // Tag 956 Vals.push_back(MDHelper.createString("VP")); 957 // Value Kind 958 Vals.push_back(MDHelper.createConstant( 959 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind))); 960 // Total Count 961 Vals.push_back( 962 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum))); 963 964 // Value Profile Data 965 uint32_t MDCount = MaxMDCount; 966 for (auto &VD : VDs) { 967 Vals.push_back(MDHelper.createConstant( 968 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value))); 969 Vals.push_back(MDHelper.createConstant( 970 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count))); 971 if (--MDCount == 0) 972 break; 973 } 974 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals)); 975 } 976 977 bool getValueProfDataFromInst(const Instruction &Inst, 978 InstrProfValueKind ValueKind, 979 uint32_t MaxNumValueData, 980 InstrProfValueData ValueData[], 981 uint32_t &ActualNumValueData, uint64_t &TotalC) { 982 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof); 983 if (!MD) 984 return false; 985 986 unsigned NOps = MD->getNumOperands(); 987 988 if (NOps < 5) 989 return false; 990 991 // Operand 0 is a string tag "VP": 992 MDString *Tag = cast<MDString>(MD->getOperand(0)); 993 if (!Tag) 994 return false; 995 996 if (!Tag->getString().equals("VP")) 997 return false; 998 999 // Now check kind: 1000 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1)); 1001 if (!KindInt) 1002 return false; 1003 if (KindInt->getZExtValue() != ValueKind) 1004 return false; 1005 1006 // Get total count 1007 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2)); 1008 if (!TotalCInt) 1009 return false; 1010 TotalC = TotalCInt->getZExtValue(); 1011 1012 ActualNumValueData = 0; 1013 1014 for (unsigned I = 3; I < NOps; I += 2) { 1015 if (ActualNumValueData >= MaxNumValueData) 1016 break; 1017 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I)); 1018 ConstantInt *Count = 1019 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1)); 1020 if (!Value || !Count) 1021 return false; 1022 ValueData[ActualNumValueData].Value = Value->getZExtValue(); 1023 ValueData[ActualNumValueData].Count = Count->getZExtValue(); 1024 ActualNumValueData++; 1025 } 1026 return true; 1027 } 1028 1029 MDNode *getPGOFuncNameMetadata(const Function &F) { 1030 return F.getMetadata(getPGOFuncNameMetadataName()); 1031 } 1032 1033 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) { 1034 // Only for internal linkage functions. 1035 if (PGOFuncName == F.getName()) 1036 return; 1037 // Don't create duplicated meta-data. 1038 if (getPGOFuncNameMetadata(F)) 1039 return; 1040 LLVMContext &C = F.getContext(); 1041 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName)); 1042 F.setMetadata(getPGOFuncNameMetadataName(), N); 1043 } 1044 1045 bool needsComdatForCounter(const Function &F, const Module &M) { 1046 if (F.hasComdat()) 1047 return true; 1048 1049 if (!Triple(M.getTargetTriple()).supportsCOMDAT()) 1050 return false; 1051 1052 // See createPGOFuncNameVar for more details. To avoid link errors, profile 1053 // counters for function with available_externally linkage needs to be changed 1054 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be 1055 // created. Without using comdat, duplicate entries won't be removed by the 1056 // linker leading to increased data segement size and raw profile size. Even 1057 // worse, since the referenced counter from profile per-function data object 1058 // will be resolved to the common strong definition, the profile counts for 1059 // available_externally functions will end up being duplicated in raw profile 1060 // data. This can result in distorted profile as the counts of those dups 1061 // will be accumulated by the profile merger. 1062 GlobalValue::LinkageTypes Linkage = F.getLinkage(); 1063 if (Linkage != GlobalValue::ExternalWeakLinkage && 1064 Linkage != GlobalValue::AvailableExternallyLinkage) 1065 return false; 1066 1067 return true; 1068 } 1069 1070 // Check if INSTR_PROF_RAW_VERSION_VAR is defined. 1071 bool isIRPGOFlagSet(const Module *M) { 1072 auto IRInstrVar = 1073 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR)); 1074 if (!IRInstrVar || IRInstrVar->isDeclaration() || 1075 IRInstrVar->hasLocalLinkage()) 1076 return false; 1077 1078 // Check if the flag is set. 1079 if (!IRInstrVar->hasInitializer()) 1080 return false; 1081 1082 const Constant *InitVal = IRInstrVar->getInitializer(); 1083 if (!InitVal) 1084 return false; 1085 1086 return (dyn_cast<ConstantInt>(InitVal)->getZExtValue() & 1087 VARIANT_MASK_IR_PROF) != 0; 1088 } 1089 1090 // Check if we can safely rename this Comdat function. 1091 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) { 1092 if (F.getName().empty()) 1093 return false; 1094 if (!needsComdatForCounter(F, *(F.getParent()))) 1095 return false; 1096 // Unsafe to rename the address-taken function (which can be used in 1097 // function comparison). 1098 if (CheckAddressTaken && F.hasAddressTaken()) 1099 return false; 1100 // Only safe to do if this function may be discarded if it is not used 1101 // in the compilation unit. 1102 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage())) 1103 return false; 1104 1105 // For AvailableExternallyLinkage functions. 1106 if (!F.hasComdat()) { 1107 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage); 1108 return true; 1109 } 1110 return true; 1111 } 1112 1113 // Parse the value profile options. 1114 void getMemOPSizeRangeFromOption(StringRef MemOPSizeRange, int64_t &RangeStart, 1115 int64_t &RangeLast) { 1116 static const int64_t DefaultMemOPSizeRangeStart = 0; 1117 static const int64_t DefaultMemOPSizeRangeLast = 8; 1118 RangeStart = DefaultMemOPSizeRangeStart; 1119 RangeLast = DefaultMemOPSizeRangeLast; 1120 1121 if (!MemOPSizeRange.empty()) { 1122 auto Pos = MemOPSizeRange.find(':'); 1123 if (Pos != std::string::npos) { 1124 if (Pos > 0) 1125 MemOPSizeRange.substr(0, Pos).getAsInteger(10, RangeStart); 1126 if (Pos < MemOPSizeRange.size() - 1) 1127 MemOPSizeRange.substr(Pos + 1).getAsInteger(10, RangeLast); 1128 } else 1129 MemOPSizeRange.getAsInteger(10, RangeLast); 1130 } 1131 assert(RangeLast >= RangeStart); 1132 } 1133 1134 // Create a COMDAT variable INSTR_PROF_RAW_VERSION_VAR to make the runtime 1135 // aware this is an ir_level profile so it can set the version flag. 1136 void createIRLevelProfileFlagVar(Module &M, bool IsCS) { 1137 const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR)); 1138 Type *IntTy64 = Type::getInt64Ty(M.getContext()); 1139 uint64_t ProfileVersion = (INSTR_PROF_RAW_VERSION | VARIANT_MASK_IR_PROF); 1140 if (IsCS) 1141 ProfileVersion |= VARIANT_MASK_CSIR_PROF; 1142 auto IRLevelVersionVariable = new GlobalVariable( 1143 M, IntTy64, true, GlobalValue::WeakAnyLinkage, 1144 Constant::getIntegerValue(IntTy64, APInt(64, ProfileVersion)), VarName); 1145 IRLevelVersionVariable->setVisibility(GlobalValue::DefaultVisibility); 1146 Triple TT(M.getTargetTriple()); 1147 if (TT.supportsCOMDAT()) { 1148 IRLevelVersionVariable->setLinkage(GlobalValue::ExternalLinkage); 1149 IRLevelVersionVariable->setComdat(M.getOrInsertComdat(VarName)); 1150 } 1151 } 1152 1153 // Create the variable for the profile file name. 1154 void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput) { 1155 if (InstrProfileOutput.empty()) 1156 return; 1157 Constant *ProfileNameConst = 1158 ConstantDataArray::getString(M.getContext(), InstrProfileOutput, true); 1159 GlobalVariable *ProfileNameVar = new GlobalVariable( 1160 M, ProfileNameConst->getType(), true, GlobalValue::WeakAnyLinkage, 1161 ProfileNameConst, INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR)); 1162 Triple TT(M.getTargetTriple()); 1163 if (TT.supportsCOMDAT()) { 1164 ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage); 1165 ProfileNameVar->setComdat(M.getOrInsertComdat( 1166 StringRef(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR)))); 1167 } 1168 } 1169 1170 Error OverlapStats::accumuateCounts(const std::string &BaseFilename, 1171 const std::string &TestFilename, 1172 bool IsCS) { 1173 auto getProfileSum = [IsCS](const std::string &Filename, 1174 CountSumOrPercent &Sum) -> Error { 1175 auto ReaderOrErr = InstrProfReader::create(Filename); 1176 if (Error E = ReaderOrErr.takeError()) { 1177 return E; 1178 } 1179 auto Reader = std::move(ReaderOrErr.get()); 1180 Reader->accumuateCounts(Sum, IsCS); 1181 return Error::success(); 1182 }; 1183 auto Ret = getProfileSum(BaseFilename, Base); 1184 if (Ret) 1185 return Ret; 1186 Ret = getProfileSum(TestFilename, Test); 1187 if (Ret) 1188 return Ret; 1189 this->BaseFilename = &BaseFilename; 1190 this->TestFilename = &TestFilename; 1191 Valid = true; 1192 return Error::success(); 1193 } 1194 1195 void OverlapStats::addOneMismatch(const CountSumOrPercent &MismatchFunc) { 1196 Mismatch.NumEntries += 1; 1197 Mismatch.CountSum += MismatchFunc.CountSum / Test.CountSum; 1198 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { 1199 if (Test.ValueCounts[I] >= 1.0f) 1200 Mismatch.ValueCounts[I] += 1201 MismatchFunc.ValueCounts[I] / Test.ValueCounts[I]; 1202 } 1203 } 1204 1205 void OverlapStats::addOneUnique(const CountSumOrPercent &UniqueFunc) { 1206 Unique.NumEntries += 1; 1207 Unique.CountSum += UniqueFunc.CountSum / Test.CountSum; 1208 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { 1209 if (Test.ValueCounts[I] >= 1.0f) 1210 Unique.ValueCounts[I] += UniqueFunc.ValueCounts[I] / Test.ValueCounts[I]; 1211 } 1212 } 1213 1214 void OverlapStats::dump(raw_fd_ostream &OS) const { 1215 if (!Valid) 1216 return; 1217 1218 const char *EntryName = 1219 (Level == ProgramLevel ? "functions" : "edge counters"); 1220 if (Level == ProgramLevel) { 1221 OS << "Profile overlap infomation for base_profile: " << *BaseFilename 1222 << " and test_profile: " << *TestFilename << "\nProgram level:\n"; 1223 } else { 1224 OS << "Function level:\n" 1225 << " Function: " << FuncName << " (Hash=" << FuncHash << ")\n"; 1226 } 1227 1228 OS << " # of " << EntryName << " overlap: " << Overlap.NumEntries << "\n"; 1229 if (Mismatch.NumEntries) 1230 OS << " # of " << EntryName << " mismatch: " << Mismatch.NumEntries 1231 << "\n"; 1232 if (Unique.NumEntries) 1233 OS << " # of " << EntryName 1234 << " only in test_profile: " << Unique.NumEntries << "\n"; 1235 1236 OS << " Edge profile overlap: " << format("%.3f%%", Overlap.CountSum * 100) 1237 << "\n"; 1238 if (Mismatch.NumEntries) 1239 OS << " Mismatched count percentage (Edge): " 1240 << format("%.3f%%", Mismatch.CountSum * 100) << "\n"; 1241 if (Unique.NumEntries) 1242 OS << " Percentage of Edge profile only in test_profile: " 1243 << format("%.3f%%", Unique.CountSum * 100) << "\n"; 1244 OS << " Edge profile base count sum: " << format("%.0f", Base.CountSum) 1245 << "\n" 1246 << " Edge profile test count sum: " << format("%.0f", Test.CountSum) 1247 << "\n"; 1248 1249 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) { 1250 if (Base.ValueCounts[I] < 1.0f && Test.ValueCounts[I] < 1.0f) 1251 continue; 1252 char ProfileKindName[20]; 1253 switch (I) { 1254 case IPVK_IndirectCallTarget: 1255 strncpy(ProfileKindName, "IndirectCall", 19); 1256 break; 1257 case IPVK_MemOPSize: 1258 strncpy(ProfileKindName, "MemOP", 19); 1259 break; 1260 default: 1261 snprintf(ProfileKindName, 19, "VP[%d]", I); 1262 break; 1263 } 1264 OS << " " << ProfileKindName 1265 << " profile overlap: " << format("%.3f%%", Overlap.ValueCounts[I] * 100) 1266 << "\n"; 1267 if (Mismatch.NumEntries) 1268 OS << " Mismatched count percentage (" << ProfileKindName 1269 << "): " << format("%.3f%%", Mismatch.ValueCounts[I] * 100) << "\n"; 1270 if (Unique.NumEntries) 1271 OS << " Percentage of " << ProfileKindName 1272 << " profile only in test_profile: " 1273 << format("%.3f%%", Unique.ValueCounts[I] * 100) << "\n"; 1274 OS << " " << ProfileKindName 1275 << " profile base count sum: " << format("%.0f", Base.ValueCounts[I]) 1276 << "\n" 1277 << " " << ProfileKindName 1278 << " profile test count sum: " << format("%.0f", Test.ValueCounts[I]) 1279 << "\n"; 1280 } 1281 } 1282 1283 } // end namespace llvm 1284