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/Config/config.h"
22 #include "llvm/IR/Constant.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GlobalValue.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/Instruction.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/MDBuilder.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/Type.h"
33 #include "llvm/ProfileData/InstrProfReader.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/Compression.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/ManagedStatic.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/Path.h"
45 #include "llvm/Support/SwapByteOrder.h"
46 #include <algorithm>
47 #include <cassert>
48 #include <cstddef>
49 #include <cstdint>
50 #include <cstring>
51 #include <memory>
52 #include <string>
53 #include <system_error>
54 #include <utility>
55 #include <vector>
56 
57 using namespace llvm;
58 
59 static cl::opt<bool> StaticFuncFullModulePrefix(
60     "static-func-full-module-prefix", cl::init(true), cl::Hidden,
61     cl::desc("Use full module build paths in the profile counter names for "
62              "static functions."));
63 
64 // This option is tailored to users that have different top-level directory in
65 // profile-gen and profile-use compilation. Users need to specific the number
66 // of levels to strip. A value larger than the number of directories in the
67 // source file will strip all the directory names and only leave the basename.
68 //
69 // Note current ThinLTO module importing for the indirect-calls assumes
70 // the source directory name not being stripped. A non-zero option value here
71 // can potentially prevent some inter-module indirect-call-promotions.
72 static cl::opt<unsigned> StaticFuncStripDirNamePrefix(
73     "static-func-strip-dirname-prefix", cl::init(0), cl::Hidden,
74     cl::desc("Strip specified level of directory name from source path in "
75              "the profile counter name for static functions."));
76 
77 static std::string getInstrProfErrString(instrprof_error Err,
78                                          const std::string &ErrMsg = "") {
79   std::string Msg;
80   raw_string_ostream OS(Msg);
81 
82   switch (Err) {
83   case instrprof_error::success:
84     OS << "success";
85     break;
86   case instrprof_error::eof:
87     OS << "end of File";
88     break;
89   case instrprof_error::unrecognized_format:
90     OS << "unrecognized instrumentation profile encoding format";
91     break;
92   case instrprof_error::bad_magic:
93     OS << "invalid instrumentation profile data (bad magic)";
94     break;
95   case instrprof_error::bad_header:
96     OS << "invalid instrumentation profile data (file header is corrupt)";
97     break;
98   case instrprof_error::unsupported_version:
99     OS << "unsupported instrumentation profile format version";
100     break;
101   case instrprof_error::unsupported_hash_type:
102     OS << "unsupported instrumentation profile hash type";
103     break;
104   case instrprof_error::too_large:
105     OS << "too much profile data";
106     break;
107   case instrprof_error::truncated:
108     OS << "truncated profile data";
109     break;
110   case instrprof_error::malformed:
111     OS << "malformed instrumentation profile data";
112     break;
113   case instrprof_error::missing_debug_info_for_correlation:
114     OS << "debug info for correlation is required";
115     break;
116   case instrprof_error::unexpected_debug_info_for_correlation:
117     OS << "debug info for correlation is not necessary";
118     break;
119   case instrprof_error::unable_to_correlate_profile:
120     OS << "unable to correlate profile";
121     break;
122   case instrprof_error::invalid_prof:
123     OS << "invalid profile created. Please file a bug "
124           "at: " BUG_REPORT_URL
125           " and include the profraw files that caused this error.";
126     break;
127   case instrprof_error::unknown_function:
128     OS << "no profile data available for function";
129     break;
130   case instrprof_error::hash_mismatch:
131     OS << "function control flow change detected (hash mismatch)";
132     break;
133   case instrprof_error::count_mismatch:
134     OS << "function basic block count change detected (counter mismatch)";
135     break;
136   case instrprof_error::counter_overflow:
137     OS << "counter overflow";
138     break;
139   case instrprof_error::value_site_count_mismatch:
140     OS << "function value site count change detected (counter mismatch)";
141     break;
142   case instrprof_error::compress_failed:
143     OS << "failed to compress data (zlib)";
144     break;
145   case instrprof_error::uncompress_failed:
146     OS << "failed to uncompress data (zlib)";
147     break;
148   case instrprof_error::empty_raw_profile:
149     OS << "empty raw profile file";
150     break;
151   case instrprof_error::zlib_unavailable:
152     OS << "profile uses zlib compression but the profile reader was built "
153           "without zlib support";
154     break;
155   }
156 
157   // If optional error message is not empty, append it to the message.
158   if (!ErrMsg.empty())
159     OS << ": " << ErrMsg;
160 
161   return OS.str();
162 }
163 
164 namespace {
165 
166 // FIXME: This class is only here to support the transition to llvm::Error. It
167 // will be removed once this transition is complete. Clients should prefer to
168 // deal with the Error value directly, rather than converting to error_code.
169 class InstrProfErrorCategoryType : public std::error_category {
170   const char *name() const noexcept override { return "llvm.instrprof"; }
171 
172   std::string message(int IE) const override {
173     return getInstrProfErrString(static_cast<instrprof_error>(IE));
174   }
175 };
176 
177 } // end anonymous namespace
178 
179 static ManagedStatic<InstrProfErrorCategoryType> ErrorCategory;
180 
181 const std::error_category &llvm::instrprof_category() {
182   return *ErrorCategory;
183 }
184 
185 namespace {
186 
187 const char *InstrProfSectNameCommon[] = {
188 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix)      \
189   SectNameCommon,
190 #include "llvm/ProfileData/InstrProfData.inc"
191 };
192 
193 const char *InstrProfSectNameCoff[] = {
194 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix)      \
195   SectNameCoff,
196 #include "llvm/ProfileData/InstrProfData.inc"
197 };
198 
199 const char *InstrProfSectNamePrefix[] = {
200 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix)      \
201   Prefix,
202 #include "llvm/ProfileData/InstrProfData.inc"
203 };
204 
205 } // namespace
206 
207 namespace llvm {
208 
209 cl::opt<bool> DoInstrProfNameCompression(
210     "enable-name-compression",
211     cl::desc("Enable name/filename string compression"), cl::init(true));
212 
213 std::string getInstrProfSectionName(InstrProfSectKind IPSK,
214                                     Triple::ObjectFormatType OF,
215                                     bool AddSegmentInfo) {
216   std::string SectName;
217 
218   if (OF == Triple::MachO && AddSegmentInfo)
219     SectName = InstrProfSectNamePrefix[IPSK];
220 
221   if (OF == Triple::COFF)
222     SectName += InstrProfSectNameCoff[IPSK];
223   else
224     SectName += InstrProfSectNameCommon[IPSK];
225 
226   if (OF == Triple::MachO && IPSK == IPSK_data && AddSegmentInfo)
227     SectName += ",regular,live_support";
228 
229   return SectName;
230 }
231 
232 void SoftInstrProfErrors::addError(instrprof_error IE) {
233   if (IE == instrprof_error::success)
234     return;
235 
236   if (FirstError == instrprof_error::success)
237     FirstError = IE;
238 
239   switch (IE) {
240   case instrprof_error::hash_mismatch:
241     ++NumHashMismatches;
242     break;
243   case instrprof_error::count_mismatch:
244     ++NumCountMismatches;
245     break;
246   case instrprof_error::counter_overflow:
247     ++NumCounterOverflows;
248     break;
249   case instrprof_error::value_site_count_mismatch:
250     ++NumValueSiteCountMismatches;
251     break;
252   default:
253     llvm_unreachable("Not a soft error");
254   }
255 }
256 
257 std::string InstrProfError::message() const {
258   return getInstrProfErrString(Err, Msg);
259 }
260 
261 char InstrProfError::ID = 0;
262 
263 std::string getPGOFuncName(StringRef RawFuncName,
264                            GlobalValue::LinkageTypes Linkage,
265                            StringRef FileName,
266                            uint64_t Version LLVM_ATTRIBUTE_UNUSED) {
267   return GlobalValue::getGlobalIdentifier(RawFuncName, Linkage, FileName);
268 }
269 
270 // Strip NumPrefix level of directory name from PathNameStr. If the number of
271 // directory separators is less than NumPrefix, strip all the directories and
272 // leave base file name only.
273 static StringRef stripDirPrefix(StringRef PathNameStr, uint32_t NumPrefix) {
274   uint32_t Count = NumPrefix;
275   uint32_t Pos = 0, LastPos = 0;
276   for (auto & CI : PathNameStr) {
277     ++Pos;
278     if (llvm::sys::path::is_separator(CI)) {
279       LastPos = Pos;
280       --Count;
281     }
282     if (Count == 0)
283       break;
284   }
285   return PathNameStr.substr(LastPos);
286 }
287 
288 // Return the PGOFuncName. This function has some special handling when called
289 // in LTO optimization. The following only applies when calling in LTO passes
290 // (when \c InLTO is true): LTO's internalization privatizes many global linkage
291 // symbols. This happens after value profile annotation, but those internal
292 // linkage functions should not have a source prefix.
293 // Additionally, for ThinLTO mode, exported internal functions are promoted
294 // and renamed. We need to ensure that the original internal PGO name is
295 // used when computing the GUID that is compared against the profiled GUIDs.
296 // To differentiate compiler generated internal symbols from original ones,
297 // PGOFuncName meta data are created and attached to the original internal
298 // symbols in the value profile annotation step
299 // (PGOUseFunc::annotateIndirectCallSites). If a symbol does not have the meta
300 // data, its original linkage must be non-internal.
301 std::string getPGOFuncName(const Function &F, bool InLTO, uint64_t Version) {
302   if (!InLTO) {
303     StringRef FileName(F.getParent()->getSourceFileName());
304     uint32_t StripLevel = StaticFuncFullModulePrefix ? 0 : (uint32_t)-1;
305     if (StripLevel < StaticFuncStripDirNamePrefix)
306       StripLevel = StaticFuncStripDirNamePrefix;
307     if (StripLevel)
308       FileName = stripDirPrefix(FileName, StripLevel);
309     return getPGOFuncName(F.getName(), F.getLinkage(), FileName, Version);
310   }
311 
312   // In LTO mode (when InLTO is true), first check if there is a meta data.
313   if (MDNode *MD = getPGOFuncNameMetadata(F)) {
314     StringRef S = cast<MDString>(MD->getOperand(0))->getString();
315     return S.str();
316   }
317 
318   // If there is no meta data, the function must be a global before the value
319   // profile annotation pass. Its current linkage may be internal if it is
320   // internalized in LTO mode.
321   return getPGOFuncName(F.getName(), GlobalValue::ExternalLinkage, "");
322 }
323 
324 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName, StringRef FileName) {
325   if (FileName.empty())
326     return PGOFuncName;
327   // Drop the file name including ':'. See also getPGOFuncName.
328   if (PGOFuncName.startswith(FileName))
329     PGOFuncName = PGOFuncName.drop_front(FileName.size() + 1);
330   return PGOFuncName;
331 }
332 
333 // \p FuncName is the string used as profile lookup key for the function. A
334 // symbol is created to hold the name. Return the legalized symbol name.
335 std::string getPGOFuncNameVarName(StringRef FuncName,
336                                   GlobalValue::LinkageTypes Linkage) {
337   std::string VarName = std::string(getInstrProfNameVarPrefix());
338   VarName += FuncName;
339 
340   if (!GlobalValue::isLocalLinkage(Linkage))
341     return VarName;
342 
343   // Now fix up illegal chars in local VarName that may upset the assembler.
344   const char *InvalidChars = "-:<>/\"'";
345   size_t found = VarName.find_first_of(InvalidChars);
346   while (found != std::string::npos) {
347     VarName[found] = '_';
348     found = VarName.find_first_of(InvalidChars, found + 1);
349   }
350   return VarName;
351 }
352 
353 GlobalVariable *createPGOFuncNameVar(Module &M,
354                                      GlobalValue::LinkageTypes Linkage,
355                                      StringRef PGOFuncName) {
356   // We generally want to match the function's linkage, but available_externally
357   // and extern_weak both have the wrong semantics, and anything that doesn't
358   // need to link across compilation units doesn't need to be visible at all.
359   if (Linkage == GlobalValue::ExternalWeakLinkage)
360     Linkage = GlobalValue::LinkOnceAnyLinkage;
361   else if (Linkage == GlobalValue::AvailableExternallyLinkage)
362     Linkage = GlobalValue::LinkOnceODRLinkage;
363   else if (Linkage == GlobalValue::InternalLinkage ||
364            Linkage == GlobalValue::ExternalLinkage)
365     Linkage = GlobalValue::PrivateLinkage;
366 
367   auto *Value =
368       ConstantDataArray::getString(M.getContext(), PGOFuncName, false);
369   auto FuncNameVar =
370       new GlobalVariable(M, Value->getType(), true, Linkage, Value,
371                          getPGOFuncNameVarName(PGOFuncName, Linkage));
372 
373   // Hide the symbol so that we correctly get a copy for each executable.
374   if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
375     FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
376 
377   return FuncNameVar;
378 }
379 
380 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
381   return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
382 }
383 
384 Error InstrProfSymtab::create(Module &M, bool InLTO) {
385   for (Function &F : M) {
386     // Function may not have a name: like using asm("") to overwrite the name.
387     // Ignore in this case.
388     if (!F.hasName())
389       continue;
390     const std::string &PGOFuncName = getPGOFuncName(F, InLTO);
391     if (Error E = addFuncName(PGOFuncName))
392       return E;
393     MD5FuncMap.emplace_back(Function::getGUID(PGOFuncName), &F);
394     // In ThinLTO, local function may have been promoted to global and have
395     // suffix ".llvm." added to the function name. We need to add the
396     // stripped function name to the symbol table so that we can find a match
397     // from profile.
398     //
399     // We may have other suffixes similar as ".llvm." which are needed to
400     // be stripped before the matching, but ".__uniq." suffix which is used
401     // to differentiate internal linkage functions in different modules
402     // should be kept. Now this is the only suffix with the pattern ".xxx"
403     // which is kept before matching.
404     const std::string UniqSuffix = ".__uniq.";
405     auto pos = PGOFuncName.find(UniqSuffix);
406     // Search '.' after ".__uniq." if ".__uniq." exists, otherwise
407     // search '.' from the beginning.
408     if (pos != std::string::npos)
409       pos += UniqSuffix.length();
410     else
411       pos = 0;
412     pos = PGOFuncName.find('.', pos);
413     if (pos != std::string::npos && pos != 0) {
414       const std::string &OtherFuncName = PGOFuncName.substr(0, pos);
415       if (Error E = addFuncName(OtherFuncName))
416         return E;
417       MD5FuncMap.emplace_back(Function::getGUID(OtherFuncName), &F);
418     }
419   }
420   Sorted = false;
421   finalizeSymtab();
422   return Error::success();
423 }
424 
425 uint64_t InstrProfSymtab::getFunctionHashFromAddress(uint64_t Address) {
426   finalizeSymtab();
427   auto It = partition_point(AddrToMD5Map, [=](std::pair<uint64_t, uint64_t> A) {
428     return A.first < Address;
429   });
430   // Raw function pointer collected by value profiler may be from
431   // external functions that are not instrumented. They won't have
432   // mapping data to be used by the deserializer. Force the value to
433   // be 0 in this case.
434   if (It != AddrToMD5Map.end() && It->first == Address)
435     return (uint64_t)It->second;
436   return 0;
437 }
438 
439 Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs,
440                                 bool doCompression, std::string &Result) {
441   assert(!NameStrs.empty() && "No name data to emit");
442 
443   uint8_t Header[16], *P = Header;
444   std::string UncompressedNameStrings =
445       join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
446 
447   assert(StringRef(UncompressedNameStrings)
448                  .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
449          "PGO name is invalid (contains separator token)");
450 
451   unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
452   P += EncLen;
453 
454   auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
455     EncLen = encodeULEB128(CompressedLen, P);
456     P += EncLen;
457     char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
458     unsigned HeaderLen = P - &Header[0];
459     Result.append(HeaderStr, HeaderLen);
460     Result += InputStr;
461     return Error::success();
462   };
463 
464   if (!doCompression) {
465     return WriteStringToResult(0, UncompressedNameStrings);
466   }
467 
468   SmallString<128> CompressedNameStrings;
469   Error E = zlib::compress(StringRef(UncompressedNameStrings),
470                            CompressedNameStrings, zlib::BestSizeCompression);
471   if (E) {
472     consumeError(std::move(E));
473     return make_error<InstrProfError>(instrprof_error::compress_failed);
474   }
475 
476   return WriteStringToResult(CompressedNameStrings.size(),
477                              CompressedNameStrings);
478 }
479 
480 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
481   auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
482   StringRef NameStr =
483       Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
484   return NameStr;
485 }
486 
487 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars,
488                                 std::string &Result, bool doCompression) {
489   std::vector<std::string> NameStrs;
490   for (auto *NameVar : NameVars) {
491     NameStrs.push_back(std::string(getPGOFuncNameVarInitializer(NameVar)));
492   }
493   return collectPGOFuncNameStrings(
494       NameStrs, zlib::isAvailable() && doCompression, Result);
495 }
496 
497 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
498   const uint8_t *P = NameStrings.bytes_begin();
499   const uint8_t *EndP = NameStrings.bytes_end();
500   while (P < EndP) {
501     uint32_t N;
502     uint64_t UncompressedSize = decodeULEB128(P, &N);
503     P += N;
504     uint64_t CompressedSize = decodeULEB128(P, &N);
505     P += N;
506     bool isCompressed = (CompressedSize != 0);
507     SmallString<128> UncompressedNameStrings;
508     StringRef NameStrings;
509     if (isCompressed) {
510       if (!llvm::zlib::isAvailable())
511         return make_error<InstrProfError>(instrprof_error::zlib_unavailable);
512 
513       StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
514                                       CompressedSize);
515       if (Error E =
516               zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
517                                UncompressedSize)) {
518         consumeError(std::move(E));
519         return make_error<InstrProfError>(instrprof_error::uncompress_failed);
520       }
521       P += CompressedSize;
522       NameStrings = StringRef(UncompressedNameStrings.data(),
523                               UncompressedNameStrings.size());
524     } else {
525       NameStrings =
526           StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
527       P += UncompressedSize;
528     }
529     // Now parse the name strings.
530     SmallVector<StringRef, 0> Names;
531     NameStrings.split(Names, getInstrProfNameSeparator());
532     for (StringRef &Name : Names)
533       if (Error E = Symtab.addFuncName(Name))
534         return E;
535 
536     while (P < EndP && *P == 0)
537       P++;
538   }
539   return Error::success();
540 }
541 
542 void InstrProfRecord::accumulateCounts(CountSumOrPercent &Sum) const {
543   uint64_t FuncSum = 0;
544   Sum.NumEntries += Counts.size();
545   for (uint64_t Count : Counts)
546     FuncSum += Count;
547   Sum.CountSum += FuncSum;
548 
549   for (uint32_t VK = IPVK_First; VK <= IPVK_Last; ++VK) {
550     uint64_t KindSum = 0;
551     uint32_t NumValueSites = getNumValueSites(VK);
552     for (size_t I = 0; I < NumValueSites; ++I) {
553       uint32_t NV = getNumValueDataForSite(VK, I);
554       std::unique_ptr<InstrProfValueData[]> VD = getValueForSite(VK, I);
555       for (uint32_t V = 0; V < NV; V++)
556         KindSum += VD[V].Count;
557     }
558     Sum.ValueCounts[VK] += KindSum;
559   }
560 }
561 
562 void InstrProfValueSiteRecord::overlap(InstrProfValueSiteRecord &Input,
563                                        uint32_t ValueKind,
564                                        OverlapStats &Overlap,
565                                        OverlapStats &FuncLevelOverlap) {
566   this->sortByTargetValues();
567   Input.sortByTargetValues();
568   double Score = 0.0f, FuncLevelScore = 0.0f;
569   auto I = ValueData.begin();
570   auto IE = ValueData.end();
571   auto J = Input.ValueData.begin();
572   auto JE = Input.ValueData.end();
573   while (I != IE && J != JE) {
574     if (I->Value == J->Value) {
575       Score += OverlapStats::score(I->Count, J->Count,
576                                    Overlap.Base.ValueCounts[ValueKind],
577                                    Overlap.Test.ValueCounts[ValueKind]);
578       FuncLevelScore += OverlapStats::score(
579           I->Count, J->Count, FuncLevelOverlap.Base.ValueCounts[ValueKind],
580           FuncLevelOverlap.Test.ValueCounts[ValueKind]);
581       ++I;
582     } else if (I->Value < J->Value) {
583       ++I;
584       continue;
585     }
586     ++J;
587   }
588   Overlap.Overlap.ValueCounts[ValueKind] += Score;
589   FuncLevelOverlap.Overlap.ValueCounts[ValueKind] += FuncLevelScore;
590 }
591 
592 // Return false on mismatch.
593 void InstrProfRecord::overlapValueProfData(uint32_t ValueKind,
594                                            InstrProfRecord &Other,
595                                            OverlapStats &Overlap,
596                                            OverlapStats &FuncLevelOverlap) {
597   uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
598   assert(ThisNumValueSites == Other.getNumValueSites(ValueKind));
599   if (!ThisNumValueSites)
600     return;
601 
602   std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
603       getOrCreateValueSitesForKind(ValueKind);
604   MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords =
605       Other.getValueSitesForKind(ValueKind);
606   for (uint32_t I = 0; I < ThisNumValueSites; I++)
607     ThisSiteRecords[I].overlap(OtherSiteRecords[I], ValueKind, Overlap,
608                                FuncLevelOverlap);
609 }
610 
611 void InstrProfRecord::overlap(InstrProfRecord &Other, OverlapStats &Overlap,
612                               OverlapStats &FuncLevelOverlap,
613                               uint64_t ValueCutoff) {
614   // FuncLevel CountSum for other should already computed and nonzero.
615   assert(FuncLevelOverlap.Test.CountSum >= 1.0f);
616   accumulateCounts(FuncLevelOverlap.Base);
617   bool Mismatch = (Counts.size() != Other.Counts.size());
618 
619   // Check if the value profiles mismatch.
620   if (!Mismatch) {
621     for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
622       uint32_t ThisNumValueSites = getNumValueSites(Kind);
623       uint32_t OtherNumValueSites = Other.getNumValueSites(Kind);
624       if (ThisNumValueSites != OtherNumValueSites) {
625         Mismatch = true;
626         break;
627       }
628     }
629   }
630   if (Mismatch) {
631     Overlap.addOneMismatch(FuncLevelOverlap.Test);
632     return;
633   }
634 
635   // Compute overlap for value counts.
636   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
637     overlapValueProfData(Kind, Other, Overlap, FuncLevelOverlap);
638 
639   double Score = 0.0;
640   uint64_t MaxCount = 0;
641   // Compute overlap for edge counts.
642   for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
643     Score += OverlapStats::score(Counts[I], Other.Counts[I],
644                                  Overlap.Base.CountSum, Overlap.Test.CountSum);
645     MaxCount = std::max(Other.Counts[I], MaxCount);
646   }
647   Overlap.Overlap.CountSum += Score;
648   Overlap.Overlap.NumEntries += 1;
649 
650   if (MaxCount >= ValueCutoff) {
651     double FuncScore = 0.0;
652     for (size_t I = 0, E = Other.Counts.size(); I < E; ++I)
653       FuncScore += OverlapStats::score(Counts[I], Other.Counts[I],
654                                        FuncLevelOverlap.Base.CountSum,
655                                        FuncLevelOverlap.Test.CountSum);
656     FuncLevelOverlap.Overlap.CountSum = FuncScore;
657     FuncLevelOverlap.Overlap.NumEntries = Other.Counts.size();
658     FuncLevelOverlap.Valid = true;
659   }
660 }
661 
662 void InstrProfValueSiteRecord::merge(InstrProfValueSiteRecord &Input,
663                                      uint64_t Weight,
664                                      function_ref<void(instrprof_error)> Warn) {
665   this->sortByTargetValues();
666   Input.sortByTargetValues();
667   auto I = ValueData.begin();
668   auto IE = ValueData.end();
669   for (const InstrProfValueData &J : Input.ValueData) {
670     while (I != IE && I->Value < J.Value)
671       ++I;
672     if (I != IE && I->Value == J.Value) {
673       bool Overflowed;
674       I->Count = SaturatingMultiplyAdd(J.Count, Weight, I->Count, &Overflowed);
675       if (Overflowed)
676         Warn(instrprof_error::counter_overflow);
677       ++I;
678       continue;
679     }
680     ValueData.insert(I, J);
681   }
682 }
683 
684 void InstrProfValueSiteRecord::scale(uint64_t N, uint64_t D,
685                                      function_ref<void(instrprof_error)> Warn) {
686   for (InstrProfValueData &I : ValueData) {
687     bool Overflowed;
688     I.Count = SaturatingMultiply(I.Count, N, &Overflowed) / D;
689     if (Overflowed)
690       Warn(instrprof_error::counter_overflow);
691   }
692 }
693 
694 // Merge Value Profile data from Src record to this record for ValueKind.
695 // Scale merged value counts by \p Weight.
696 void InstrProfRecord::mergeValueProfData(
697     uint32_t ValueKind, InstrProfRecord &Src, uint64_t Weight,
698     function_ref<void(instrprof_error)> Warn) {
699   uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
700   uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
701   if (ThisNumValueSites != OtherNumValueSites) {
702     Warn(instrprof_error::value_site_count_mismatch);
703     return;
704   }
705   if (!ThisNumValueSites)
706     return;
707   std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
708       getOrCreateValueSitesForKind(ValueKind);
709   MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords =
710       Src.getValueSitesForKind(ValueKind);
711   for (uint32_t I = 0; I < ThisNumValueSites; I++)
712     ThisSiteRecords[I].merge(OtherSiteRecords[I], Weight, Warn);
713 }
714 
715 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight,
716                             function_ref<void(instrprof_error)> Warn) {
717   // If the number of counters doesn't match we either have bad data
718   // or a hash collision.
719   if (Counts.size() != Other.Counts.size()) {
720     Warn(instrprof_error::count_mismatch);
721     return;
722   }
723 
724   for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
725     bool Overflowed;
726     Counts[I] =
727         SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
728     if (Overflowed)
729       Warn(instrprof_error::counter_overflow);
730   }
731 
732   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
733     mergeValueProfData(Kind, Other, Weight, Warn);
734 }
735 
736 void InstrProfRecord::scaleValueProfData(
737     uint32_t ValueKind, uint64_t N, uint64_t D,
738     function_ref<void(instrprof_error)> Warn) {
739   for (auto &R : getValueSitesForKind(ValueKind))
740     R.scale(N, D, Warn);
741 }
742 
743 void InstrProfRecord::scale(uint64_t N, uint64_t D,
744                             function_ref<void(instrprof_error)> Warn) {
745   assert(D != 0 && "D cannot be 0");
746   for (auto &Count : this->Counts) {
747     bool Overflowed;
748     Count = SaturatingMultiply(Count, N, &Overflowed) / D;
749     if (Overflowed)
750       Warn(instrprof_error::counter_overflow);
751   }
752   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
753     scaleValueProfData(Kind, N, D, Warn);
754 }
755 
756 // Map indirect call target name hash to name string.
757 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
758                                      InstrProfSymtab *SymTab) {
759   if (!SymTab)
760     return Value;
761 
762   if (ValueKind == IPVK_IndirectCallTarget)
763     return SymTab->getFunctionHashFromAddress(Value);
764 
765   return Value;
766 }
767 
768 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
769                                    InstrProfValueData *VData, uint32_t N,
770                                    InstrProfSymtab *ValueMap) {
771   for (uint32_t I = 0; I < N; I++) {
772     VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
773   }
774   std::vector<InstrProfValueSiteRecord> &ValueSites =
775       getOrCreateValueSitesForKind(ValueKind);
776   if (N == 0)
777     ValueSites.emplace_back();
778   else
779     ValueSites.emplace_back(VData, VData + N);
780 }
781 
782 #define INSTR_PROF_COMMON_API_IMPL
783 #include "llvm/ProfileData/InstrProfData.inc"
784 
785 /*!
786  * ValueProfRecordClosure Interface implementation for  InstrProfRecord
787  *  class. These C wrappers are used as adaptors so that C++ code can be
788  *  invoked as callbacks.
789  */
790 uint32_t getNumValueKindsInstrProf(const void *Record) {
791   return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
792 }
793 
794 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
795   return reinterpret_cast<const InstrProfRecord *>(Record)
796       ->getNumValueSites(VKind);
797 }
798 
799 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
800   return reinterpret_cast<const InstrProfRecord *>(Record)
801       ->getNumValueData(VKind);
802 }
803 
804 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
805                                          uint32_t S) {
806   return reinterpret_cast<const InstrProfRecord *>(R)
807       ->getNumValueDataForSite(VK, S);
808 }
809 
810 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
811                               uint32_t K, uint32_t S) {
812   reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
813 }
814 
815 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
816   ValueProfData *VD =
817       (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
818   memset(VD, 0, TotalSizeInBytes);
819   return VD;
820 }
821 
822 static ValueProfRecordClosure InstrProfRecordClosure = {
823     nullptr,
824     getNumValueKindsInstrProf,
825     getNumValueSitesInstrProf,
826     getNumValueDataInstrProf,
827     getNumValueDataForSiteInstrProf,
828     nullptr,
829     getValueForSiteInstrProf,
830     allocValueProfDataInstrProf};
831 
832 // Wrapper implementation using the closure mechanism.
833 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
834   auto Closure = InstrProfRecordClosure;
835   Closure.Record = &Record;
836   return getValueProfDataSize(&Closure);
837 }
838 
839 // Wrapper implementation using the closure mechanism.
840 std::unique_ptr<ValueProfData>
841 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
842   InstrProfRecordClosure.Record = &Record;
843 
844   std::unique_ptr<ValueProfData> VPD(
845       serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
846   return VPD;
847 }
848 
849 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
850                                     InstrProfSymtab *SymTab) {
851   Record.reserveSites(Kind, NumValueSites);
852 
853   InstrProfValueData *ValueData = getValueProfRecordValueData(this);
854   for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
855     uint8_t ValueDataCount = this->SiteCountArray[VSite];
856     Record.addValueData(Kind, VSite, ValueData, ValueDataCount, SymTab);
857     ValueData += ValueDataCount;
858   }
859 }
860 
861 // For writing/serializing,  Old is the host endianness, and  New is
862 // byte order intended on disk. For Reading/deserialization, Old
863 // is the on-disk source endianness, and New is the host endianness.
864 void ValueProfRecord::swapBytes(support::endianness Old,
865                                 support::endianness New) {
866   using namespace support;
867 
868   if (Old == New)
869     return;
870 
871   if (getHostEndianness() != Old) {
872     sys::swapByteOrder<uint32_t>(NumValueSites);
873     sys::swapByteOrder<uint32_t>(Kind);
874   }
875   uint32_t ND = getValueProfRecordNumValueData(this);
876   InstrProfValueData *VD = getValueProfRecordValueData(this);
877 
878   // No need to swap byte array: SiteCountArrray.
879   for (uint32_t I = 0; I < ND; I++) {
880     sys::swapByteOrder<uint64_t>(VD[I].Value);
881     sys::swapByteOrder<uint64_t>(VD[I].Count);
882   }
883   if (getHostEndianness() == Old) {
884     sys::swapByteOrder<uint32_t>(NumValueSites);
885     sys::swapByteOrder<uint32_t>(Kind);
886   }
887 }
888 
889 void ValueProfData::deserializeTo(InstrProfRecord &Record,
890                                   InstrProfSymtab *SymTab) {
891   if (NumValueKinds == 0)
892     return;
893 
894   ValueProfRecord *VR = getFirstValueProfRecord(this);
895   for (uint32_t K = 0; K < NumValueKinds; K++) {
896     VR->deserializeTo(Record, SymTab);
897     VR = getValueProfRecordNext(VR);
898   }
899 }
900 
901 template <class T>
902 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
903   using namespace support;
904 
905   if (Orig == little)
906     return endian::readNext<T, little, unaligned>(D);
907   else
908     return endian::readNext<T, big, unaligned>(D);
909 }
910 
911 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
912   return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
913                                             ValueProfData());
914 }
915 
916 Error ValueProfData::checkIntegrity() {
917   if (NumValueKinds > IPVK_Last + 1)
918     return make_error<InstrProfError>(
919         instrprof_error::malformed, "number of value profile kinds is invalid");
920   // Total size needs to be multiple of quadword size.
921   if (TotalSize % sizeof(uint64_t))
922     return make_error<InstrProfError>(
923         instrprof_error::malformed, "total size is not multiples of quardword");
924 
925   ValueProfRecord *VR = getFirstValueProfRecord(this);
926   for (uint32_t K = 0; K < this->NumValueKinds; K++) {
927     if (VR->Kind > IPVK_Last)
928       return make_error<InstrProfError>(instrprof_error::malformed,
929                                         "value kind is invalid");
930     VR = getValueProfRecordNext(VR);
931     if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
932       return make_error<InstrProfError>(
933           instrprof_error::malformed,
934           "value profile address is greater than total size");
935   }
936   return Error::success();
937 }
938 
939 Expected<std::unique_ptr<ValueProfData>>
940 ValueProfData::getValueProfData(const unsigned char *D,
941                                 const unsigned char *const BufferEnd,
942                                 support::endianness Endianness) {
943   using namespace support;
944 
945   if (D + sizeof(ValueProfData) > BufferEnd)
946     return make_error<InstrProfError>(instrprof_error::truncated);
947 
948   const unsigned char *Header = D;
949   uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
950   if (D + TotalSize > BufferEnd)
951     return make_error<InstrProfError>(instrprof_error::too_large);
952 
953   std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
954   memcpy(VPD.get(), D, TotalSize);
955   // Byte swap.
956   VPD->swapBytesToHost(Endianness);
957 
958   Error E = VPD->checkIntegrity();
959   if (E)
960     return std::move(E);
961 
962   return std::move(VPD);
963 }
964 
965 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
966   using namespace support;
967 
968   if (Endianness == getHostEndianness())
969     return;
970 
971   sys::swapByteOrder<uint32_t>(TotalSize);
972   sys::swapByteOrder<uint32_t>(NumValueKinds);
973 
974   ValueProfRecord *VR = getFirstValueProfRecord(this);
975   for (uint32_t K = 0; K < NumValueKinds; K++) {
976     VR->swapBytes(Endianness, getHostEndianness());
977     VR = getValueProfRecordNext(VR);
978   }
979 }
980 
981 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
982   using namespace support;
983 
984   if (Endianness == getHostEndianness())
985     return;
986 
987   ValueProfRecord *VR = getFirstValueProfRecord(this);
988   for (uint32_t K = 0; K < NumValueKinds; K++) {
989     ValueProfRecord *NVR = getValueProfRecordNext(VR);
990     VR->swapBytes(getHostEndianness(), Endianness);
991     VR = NVR;
992   }
993   sys::swapByteOrder<uint32_t>(TotalSize);
994   sys::swapByteOrder<uint32_t>(NumValueKinds);
995 }
996 
997 void annotateValueSite(Module &M, Instruction &Inst,
998                        const InstrProfRecord &InstrProfR,
999                        InstrProfValueKind ValueKind, uint32_t SiteIdx,
1000                        uint32_t MaxMDCount) {
1001   uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
1002   if (!NV)
1003     return;
1004 
1005   uint64_t Sum = 0;
1006   std::unique_ptr<InstrProfValueData[]> VD =
1007       InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
1008 
1009   ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
1010   annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
1011 }
1012 
1013 void annotateValueSite(Module &M, Instruction &Inst,
1014                        ArrayRef<InstrProfValueData> VDs,
1015                        uint64_t Sum, InstrProfValueKind ValueKind,
1016                        uint32_t MaxMDCount) {
1017   LLVMContext &Ctx = M.getContext();
1018   MDBuilder MDHelper(Ctx);
1019   SmallVector<Metadata *, 3> Vals;
1020   // Tag
1021   Vals.push_back(MDHelper.createString("VP"));
1022   // Value Kind
1023   Vals.push_back(MDHelper.createConstant(
1024       ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
1025   // Total Count
1026   Vals.push_back(
1027       MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
1028 
1029   // Value Profile Data
1030   uint32_t MDCount = MaxMDCount;
1031   for (auto &VD : VDs) {
1032     Vals.push_back(MDHelper.createConstant(
1033         ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
1034     Vals.push_back(MDHelper.createConstant(
1035         ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
1036     if (--MDCount == 0)
1037       break;
1038   }
1039   Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
1040 }
1041 
1042 bool getValueProfDataFromInst(const Instruction &Inst,
1043                               InstrProfValueKind ValueKind,
1044                               uint32_t MaxNumValueData,
1045                               InstrProfValueData ValueData[],
1046                               uint32_t &ActualNumValueData, uint64_t &TotalC,
1047                               bool GetNoICPValue) {
1048   MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
1049   if (!MD)
1050     return false;
1051 
1052   unsigned NOps = MD->getNumOperands();
1053 
1054   if (NOps < 5)
1055     return false;
1056 
1057   // Operand 0 is a string tag "VP":
1058   MDString *Tag = cast<MDString>(MD->getOperand(0));
1059   if (!Tag)
1060     return false;
1061 
1062   if (!Tag->getString().equals("VP"))
1063     return false;
1064 
1065   // Now check kind:
1066   ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
1067   if (!KindInt)
1068     return false;
1069   if (KindInt->getZExtValue() != ValueKind)
1070     return false;
1071 
1072   // Get total count
1073   ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
1074   if (!TotalCInt)
1075     return false;
1076   TotalC = TotalCInt->getZExtValue();
1077 
1078   ActualNumValueData = 0;
1079 
1080   for (unsigned I = 3; I < NOps; I += 2) {
1081     if (ActualNumValueData >= MaxNumValueData)
1082       break;
1083     ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
1084     ConstantInt *Count =
1085         mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
1086     if (!Value || !Count)
1087       return false;
1088     uint64_t CntValue = Count->getZExtValue();
1089     if (!GetNoICPValue && (CntValue == NOMORE_ICP_MAGICNUM))
1090       continue;
1091     ValueData[ActualNumValueData].Value = Value->getZExtValue();
1092     ValueData[ActualNumValueData].Count = CntValue;
1093     ActualNumValueData++;
1094   }
1095   return true;
1096 }
1097 
1098 MDNode *getPGOFuncNameMetadata(const Function &F) {
1099   return F.getMetadata(getPGOFuncNameMetadataName());
1100 }
1101 
1102 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
1103   // Only for internal linkage functions.
1104   if (PGOFuncName == F.getName())
1105       return;
1106   // Don't create duplicated meta-data.
1107   if (getPGOFuncNameMetadata(F))
1108     return;
1109   LLVMContext &C = F.getContext();
1110   MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
1111   F.setMetadata(getPGOFuncNameMetadataName(), N);
1112 }
1113 
1114 bool needsComdatForCounter(const Function &F, const Module &M) {
1115   if (F.hasComdat())
1116     return true;
1117 
1118   if (!Triple(M.getTargetTriple()).supportsCOMDAT())
1119     return false;
1120 
1121   // See createPGOFuncNameVar for more details. To avoid link errors, profile
1122   // counters for function with available_externally linkage needs to be changed
1123   // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
1124   // created. Without using comdat, duplicate entries won't be removed by the
1125   // linker leading to increased data segement size and raw profile size. Even
1126   // worse, since the referenced counter from profile per-function data object
1127   // will be resolved to the common strong definition, the profile counts for
1128   // available_externally functions will end up being duplicated in raw profile
1129   // data. This can result in distorted profile as the counts of those dups
1130   // will be accumulated by the profile merger.
1131   GlobalValue::LinkageTypes Linkage = F.getLinkage();
1132   if (Linkage != GlobalValue::ExternalWeakLinkage &&
1133       Linkage != GlobalValue::AvailableExternallyLinkage)
1134     return false;
1135 
1136   return true;
1137 }
1138 
1139 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
1140 bool isIRPGOFlagSet(const Module *M) {
1141   auto IRInstrVar =
1142       M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
1143   if (!IRInstrVar || IRInstrVar->hasLocalLinkage())
1144     return false;
1145 
1146   // For CSPGO+LTO, this variable might be marked as non-prevailing and we only
1147   // have the decl.
1148   if (IRInstrVar->isDeclaration())
1149     return true;
1150 
1151   // Check if the flag is set.
1152   if (!IRInstrVar->hasInitializer())
1153     return false;
1154 
1155   auto *InitVal = dyn_cast_or_null<ConstantInt>(IRInstrVar->getInitializer());
1156   if (!InitVal)
1157     return false;
1158   return (InitVal->getZExtValue() & VARIANT_MASK_IR_PROF) != 0;
1159 }
1160 
1161 // Check if we can safely rename this Comdat function.
1162 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
1163   if (F.getName().empty())
1164     return false;
1165   if (!needsComdatForCounter(F, *(F.getParent())))
1166     return false;
1167   // Unsafe to rename the address-taken function (which can be used in
1168   // function comparison).
1169   if (CheckAddressTaken && F.hasAddressTaken())
1170     return false;
1171   // Only safe to do if this function may be discarded if it is not used
1172   // in the compilation unit.
1173   if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
1174     return false;
1175 
1176   // For AvailableExternallyLinkage functions.
1177   if (!F.hasComdat()) {
1178     assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
1179     return true;
1180   }
1181   return true;
1182 }
1183 
1184 // Create the variable for the profile file name.
1185 void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput) {
1186   if (InstrProfileOutput.empty())
1187     return;
1188   Constant *ProfileNameConst =
1189       ConstantDataArray::getString(M.getContext(), InstrProfileOutput, true);
1190   GlobalVariable *ProfileNameVar = new GlobalVariable(
1191       M, ProfileNameConst->getType(), true, GlobalValue::WeakAnyLinkage,
1192       ProfileNameConst, INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR));
1193   Triple TT(M.getTargetTriple());
1194   if (TT.supportsCOMDAT()) {
1195     ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage);
1196     ProfileNameVar->setComdat(M.getOrInsertComdat(
1197         StringRef(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR))));
1198   }
1199 }
1200 
1201 Error OverlapStats::accumulateCounts(const std::string &BaseFilename,
1202                                      const std::string &TestFilename,
1203                                      bool IsCS) {
1204   auto getProfileSum = [IsCS](const std::string &Filename,
1205                               CountSumOrPercent &Sum) -> Error {
1206     auto ReaderOrErr = InstrProfReader::create(Filename);
1207     if (Error E = ReaderOrErr.takeError()) {
1208       return E;
1209     }
1210     auto Reader = std::move(ReaderOrErr.get());
1211     Reader->accumulateCounts(Sum, IsCS);
1212     return Error::success();
1213   };
1214   auto Ret = getProfileSum(BaseFilename, Base);
1215   if (Ret)
1216     return Ret;
1217   Ret = getProfileSum(TestFilename, Test);
1218   if (Ret)
1219     return Ret;
1220   this->BaseFilename = &BaseFilename;
1221   this->TestFilename = &TestFilename;
1222   Valid = true;
1223   return Error::success();
1224 }
1225 
1226 void OverlapStats::addOneMismatch(const CountSumOrPercent &MismatchFunc) {
1227   Mismatch.NumEntries += 1;
1228   Mismatch.CountSum += MismatchFunc.CountSum / Test.CountSum;
1229   for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1230     if (Test.ValueCounts[I] >= 1.0f)
1231       Mismatch.ValueCounts[I] +=
1232           MismatchFunc.ValueCounts[I] / Test.ValueCounts[I];
1233   }
1234 }
1235 
1236 void OverlapStats::addOneUnique(const CountSumOrPercent &UniqueFunc) {
1237   Unique.NumEntries += 1;
1238   Unique.CountSum += UniqueFunc.CountSum / Test.CountSum;
1239   for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1240     if (Test.ValueCounts[I] >= 1.0f)
1241       Unique.ValueCounts[I] += UniqueFunc.ValueCounts[I] / Test.ValueCounts[I];
1242   }
1243 }
1244 
1245 void OverlapStats::dump(raw_fd_ostream &OS) const {
1246   if (!Valid)
1247     return;
1248 
1249   const char *EntryName =
1250       (Level == ProgramLevel ? "functions" : "edge counters");
1251   if (Level == ProgramLevel) {
1252     OS << "Profile overlap infomation for base_profile: " << *BaseFilename
1253        << " and test_profile: " << *TestFilename << "\nProgram level:\n";
1254   } else {
1255     OS << "Function level:\n"
1256        << "  Function: " << FuncName << " (Hash=" << FuncHash << ")\n";
1257   }
1258 
1259   OS << "  # of " << EntryName << " overlap: " << Overlap.NumEntries << "\n";
1260   if (Mismatch.NumEntries)
1261     OS << "  # of " << EntryName << " mismatch: " << Mismatch.NumEntries
1262        << "\n";
1263   if (Unique.NumEntries)
1264     OS << "  # of " << EntryName
1265        << " only in test_profile: " << Unique.NumEntries << "\n";
1266 
1267   OS << "  Edge profile overlap: " << format("%.3f%%", Overlap.CountSum * 100)
1268      << "\n";
1269   if (Mismatch.NumEntries)
1270     OS << "  Mismatched count percentage (Edge): "
1271        << format("%.3f%%", Mismatch.CountSum * 100) << "\n";
1272   if (Unique.NumEntries)
1273     OS << "  Percentage of Edge profile only in test_profile: "
1274        << format("%.3f%%", Unique.CountSum * 100) << "\n";
1275   OS << "  Edge profile base count sum: " << format("%.0f", Base.CountSum)
1276      << "\n"
1277      << "  Edge profile test count sum: " << format("%.0f", Test.CountSum)
1278      << "\n";
1279 
1280   for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1281     if (Base.ValueCounts[I] < 1.0f && Test.ValueCounts[I] < 1.0f)
1282       continue;
1283     char ProfileKindName[20];
1284     switch (I) {
1285     case IPVK_IndirectCallTarget:
1286       strncpy(ProfileKindName, "IndirectCall", 19);
1287       break;
1288     case IPVK_MemOPSize:
1289       strncpy(ProfileKindName, "MemOP", 19);
1290       break;
1291     default:
1292       snprintf(ProfileKindName, 19, "VP[%d]", I);
1293       break;
1294     }
1295     OS << "  " << ProfileKindName
1296        << " profile overlap: " << format("%.3f%%", Overlap.ValueCounts[I] * 100)
1297        << "\n";
1298     if (Mismatch.NumEntries)
1299       OS << "  Mismatched count percentage (" << ProfileKindName
1300          << "): " << format("%.3f%%", Mismatch.ValueCounts[I] * 100) << "\n";
1301     if (Unique.NumEntries)
1302       OS << "  Percentage of " << ProfileKindName
1303          << " profile only in test_profile: "
1304          << format("%.3f%%", Unique.ValueCounts[I] * 100) << "\n";
1305     OS << "  " << ProfileKindName
1306        << " profile base count sum: " << format("%.0f", Base.ValueCounts[I])
1307        << "\n"
1308        << "  " << ProfileKindName
1309        << " profile test count sum: " << format("%.0f", Test.ValueCounts[I])
1310        << "\n";
1311   }
1312 }
1313 
1314 } // end namespace llvm
1315