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
getInstrProfErrString(instrprof_error Err)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 {
name() const126 const char *name() const noexcept override { return "llvm.instrprof"; }
127
message(int IE) const128 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
instrprof_category()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
getInstrProfSectionName(InstrProfSectKind IPSK,Triple::ObjectFormatType OF,bool AddSegmentInfo)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
addError(instrprof_error IE)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
message() const209 std::string InstrProfError::message() const {
210 return getInstrProfErrString(Err);
211 }
212
213 char InstrProfError::ID = 0;
214
getPGOFuncName(StringRef RawFuncName,GlobalValue::LinkageTypes Linkage,StringRef FileName,uint64_t Version LLVM_ATTRIBUTE_UNUSED)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.
stripDirPrefix(StringRef PathNameStr,uint32_t NumPrefix)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.
getPGOFuncName(const Function & F,bool InLTO,uint64_t Version)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
getFuncNameWithoutPrefix(StringRef PGOFuncName,StringRef FileName)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.
getPGOFuncNameVarName(StringRef FuncName,GlobalValue::LinkageTypes Linkage)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
createPGOFuncNameVar(Module & M,GlobalValue::LinkageTypes Linkage,StringRef PGOFuncName)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
createPGOFuncNameVar(Function & F,StringRef PGOFuncName)332 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName) {
333 return createPGOFuncNameVar(*F.getParent(), F.getLinkage(), PGOFuncName);
334 }
335
create(Module & M,bool InLTO)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
getFunctionHashFromAddress(uint64_t Address)364 uint64_t InstrProfSymtab::getFunctionHashFromAddress(uint64_t Address) {
365 finalizeSymtab();
366 auto It = partition_point(AddrToMD5Map, [=](std::pair<uint64_t, uint64_t> A) {
367 return A.first < Address;
368 });
369 // Raw function pointer collected by value profiler may be from
370 // external functions that are not instrumented. They won't have
371 // mapping data to be used by the deserializer. Force the value to
372 // be 0 in this case.
373 if (It != AddrToMD5Map.end() && It->first == Address)
374 return (uint64_t)It->second;
375 return 0;
376 }
377
collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs,bool doCompression,std::string & Result)378 Error collectPGOFuncNameStrings(ArrayRef<std::string> NameStrs,
379 bool doCompression, std::string &Result) {
380 assert(!NameStrs.empty() && "No name data to emit");
381
382 uint8_t Header[16], *P = Header;
383 std::string UncompressedNameStrings =
384 join(NameStrs.begin(), NameStrs.end(), getInstrProfNameSeparator());
385
386 assert(StringRef(UncompressedNameStrings)
387 .count(getInstrProfNameSeparator()) == (NameStrs.size() - 1) &&
388 "PGO name is invalid (contains separator token)");
389
390 unsigned EncLen = encodeULEB128(UncompressedNameStrings.length(), P);
391 P += EncLen;
392
393 auto WriteStringToResult = [&](size_t CompressedLen, StringRef InputStr) {
394 EncLen = encodeULEB128(CompressedLen, P);
395 P += EncLen;
396 char *HeaderStr = reinterpret_cast<char *>(&Header[0]);
397 unsigned HeaderLen = P - &Header[0];
398 Result.append(HeaderStr, HeaderLen);
399 Result += InputStr;
400 return Error::success();
401 };
402
403 if (!doCompression) {
404 return WriteStringToResult(0, UncompressedNameStrings);
405 }
406
407 SmallString<128> CompressedNameStrings;
408 Error E = zlib::compress(StringRef(UncompressedNameStrings),
409 CompressedNameStrings, zlib::BestSizeCompression);
410 if (E) {
411 consumeError(std::move(E));
412 return make_error<InstrProfError>(instrprof_error::compress_failed);
413 }
414
415 return WriteStringToResult(CompressedNameStrings.size(),
416 CompressedNameStrings);
417 }
418
getPGOFuncNameVarInitializer(GlobalVariable * NameVar)419 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar) {
420 auto *Arr = cast<ConstantDataArray>(NameVar->getInitializer());
421 StringRef NameStr =
422 Arr->isCString() ? Arr->getAsCString() : Arr->getAsString();
423 return NameStr;
424 }
425
collectPGOFuncNameStrings(ArrayRef<GlobalVariable * > NameVars,std::string & Result,bool doCompression)426 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars,
427 std::string &Result, bool doCompression) {
428 std::vector<std::string> NameStrs;
429 for (auto *NameVar : NameVars) {
430 NameStrs.push_back(getPGOFuncNameVarInitializer(NameVar));
431 }
432 return collectPGOFuncNameStrings(
433 NameStrs, zlib::isAvailable() && doCompression, Result);
434 }
435
readPGOFuncNameStrings(StringRef NameStrings,InstrProfSymtab & Symtab)436 Error readPGOFuncNameStrings(StringRef NameStrings, InstrProfSymtab &Symtab) {
437 const uint8_t *P = NameStrings.bytes_begin();
438 const uint8_t *EndP = NameStrings.bytes_end();
439 while (P < EndP) {
440 uint32_t N;
441 uint64_t UncompressedSize = decodeULEB128(P, &N);
442 P += N;
443 uint64_t CompressedSize = decodeULEB128(P, &N);
444 P += N;
445 bool isCompressed = (CompressedSize != 0);
446 SmallString<128> UncompressedNameStrings;
447 StringRef NameStrings;
448 if (isCompressed) {
449 if (!llvm::zlib::isAvailable())
450 return make_error<InstrProfError>(instrprof_error::zlib_unavailable);
451
452 StringRef CompressedNameStrings(reinterpret_cast<const char *>(P),
453 CompressedSize);
454 if (Error E =
455 zlib::uncompress(CompressedNameStrings, UncompressedNameStrings,
456 UncompressedSize)) {
457 consumeError(std::move(E));
458 return make_error<InstrProfError>(instrprof_error::uncompress_failed);
459 }
460 P += CompressedSize;
461 NameStrings = StringRef(UncompressedNameStrings.data(),
462 UncompressedNameStrings.size());
463 } else {
464 NameStrings =
465 StringRef(reinterpret_cast<const char *>(P), UncompressedSize);
466 P += UncompressedSize;
467 }
468 // Now parse the name strings.
469 SmallVector<StringRef, 0> Names;
470 NameStrings.split(Names, getInstrProfNameSeparator());
471 for (StringRef &Name : Names)
472 if (Error E = Symtab.addFuncName(Name))
473 return E;
474
475 while (P < EndP && *P == 0)
476 P++;
477 }
478 return Error::success();
479 }
480
accumulateCounts(CountSumOrPercent & Sum) const481 void InstrProfRecord::accumulateCounts(CountSumOrPercent &Sum) const {
482 uint64_t FuncSum = 0;
483 Sum.NumEntries += Counts.size();
484 for (size_t F = 0, E = Counts.size(); F < E; ++F)
485 FuncSum += Counts[F];
486 Sum.CountSum += FuncSum;
487
488 for (uint32_t VK = IPVK_First; VK <= IPVK_Last; ++VK) {
489 uint64_t KindSum = 0;
490 uint32_t NumValueSites = getNumValueSites(VK);
491 for (size_t I = 0; I < NumValueSites; ++I) {
492 uint32_t NV = getNumValueDataForSite(VK, I);
493 std::unique_ptr<InstrProfValueData[]> VD = getValueForSite(VK, I);
494 for (uint32_t V = 0; V < NV; V++)
495 KindSum += VD[V].Count;
496 }
497 Sum.ValueCounts[VK] += KindSum;
498 }
499 }
500
overlap(InstrProfValueSiteRecord & Input,uint32_t ValueKind,OverlapStats & Overlap,OverlapStats & FuncLevelOverlap)501 void InstrProfValueSiteRecord::overlap(InstrProfValueSiteRecord &Input,
502 uint32_t ValueKind,
503 OverlapStats &Overlap,
504 OverlapStats &FuncLevelOverlap) {
505 this->sortByTargetValues();
506 Input.sortByTargetValues();
507 double Score = 0.0f, FuncLevelScore = 0.0f;
508 auto I = ValueData.begin();
509 auto IE = ValueData.end();
510 auto J = Input.ValueData.begin();
511 auto JE = Input.ValueData.end();
512 while (I != IE && J != JE) {
513 if (I->Value == J->Value) {
514 Score += OverlapStats::score(I->Count, J->Count,
515 Overlap.Base.ValueCounts[ValueKind],
516 Overlap.Test.ValueCounts[ValueKind]);
517 FuncLevelScore += OverlapStats::score(
518 I->Count, J->Count, FuncLevelOverlap.Base.ValueCounts[ValueKind],
519 FuncLevelOverlap.Test.ValueCounts[ValueKind]);
520 ++I;
521 } else if (I->Value < J->Value) {
522 ++I;
523 continue;
524 }
525 ++J;
526 }
527 Overlap.Overlap.ValueCounts[ValueKind] += Score;
528 FuncLevelOverlap.Overlap.ValueCounts[ValueKind] += FuncLevelScore;
529 }
530
531 // Return false on mismatch.
overlapValueProfData(uint32_t ValueKind,InstrProfRecord & Other,OverlapStats & Overlap,OverlapStats & FuncLevelOverlap)532 void InstrProfRecord::overlapValueProfData(uint32_t ValueKind,
533 InstrProfRecord &Other,
534 OverlapStats &Overlap,
535 OverlapStats &FuncLevelOverlap) {
536 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
537 assert(ThisNumValueSites == Other.getNumValueSites(ValueKind));
538 if (!ThisNumValueSites)
539 return;
540
541 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
542 getOrCreateValueSitesForKind(ValueKind);
543 MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords =
544 Other.getValueSitesForKind(ValueKind);
545 for (uint32_t I = 0; I < ThisNumValueSites; I++)
546 ThisSiteRecords[I].overlap(OtherSiteRecords[I], ValueKind, Overlap,
547 FuncLevelOverlap);
548 }
549
overlap(InstrProfRecord & Other,OverlapStats & Overlap,OverlapStats & FuncLevelOverlap,uint64_t ValueCutoff)550 void InstrProfRecord::overlap(InstrProfRecord &Other, OverlapStats &Overlap,
551 OverlapStats &FuncLevelOverlap,
552 uint64_t ValueCutoff) {
553 // FuncLevel CountSum for other should already computed and nonzero.
554 assert(FuncLevelOverlap.Test.CountSum >= 1.0f);
555 accumulateCounts(FuncLevelOverlap.Base);
556 bool Mismatch = (Counts.size() != Other.Counts.size());
557
558 // Check if the value profiles mismatch.
559 if (!Mismatch) {
560 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
561 uint32_t ThisNumValueSites = getNumValueSites(Kind);
562 uint32_t OtherNumValueSites = Other.getNumValueSites(Kind);
563 if (ThisNumValueSites != OtherNumValueSites) {
564 Mismatch = true;
565 break;
566 }
567 }
568 }
569 if (Mismatch) {
570 Overlap.addOneMismatch(FuncLevelOverlap.Test);
571 return;
572 }
573
574 // Compute overlap for value counts.
575 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
576 overlapValueProfData(Kind, Other, Overlap, FuncLevelOverlap);
577
578 double Score = 0.0;
579 uint64_t MaxCount = 0;
580 // Compute overlap for edge counts.
581 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
582 Score += OverlapStats::score(Counts[I], Other.Counts[I],
583 Overlap.Base.CountSum, Overlap.Test.CountSum);
584 MaxCount = std::max(Other.Counts[I], MaxCount);
585 }
586 Overlap.Overlap.CountSum += Score;
587 Overlap.Overlap.NumEntries += 1;
588
589 if (MaxCount >= ValueCutoff) {
590 double FuncScore = 0.0;
591 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I)
592 FuncScore += OverlapStats::score(Counts[I], Other.Counts[I],
593 FuncLevelOverlap.Base.CountSum,
594 FuncLevelOverlap.Test.CountSum);
595 FuncLevelOverlap.Overlap.CountSum = FuncScore;
596 FuncLevelOverlap.Overlap.NumEntries = Other.Counts.size();
597 FuncLevelOverlap.Valid = true;
598 }
599 }
600
merge(InstrProfValueSiteRecord & Input,uint64_t Weight,function_ref<void (instrprof_error)> Warn)601 void InstrProfValueSiteRecord::merge(InstrProfValueSiteRecord &Input,
602 uint64_t Weight,
603 function_ref<void(instrprof_error)> Warn) {
604 this->sortByTargetValues();
605 Input.sortByTargetValues();
606 auto I = ValueData.begin();
607 auto IE = ValueData.end();
608 for (auto J = Input.ValueData.begin(), JE = Input.ValueData.end(); J != JE;
609 ++J) {
610 while (I != IE && I->Value < J->Value)
611 ++I;
612 if (I != IE && I->Value == J->Value) {
613 bool Overflowed;
614 I->Count = SaturatingMultiplyAdd(J->Count, Weight, I->Count, &Overflowed);
615 if (Overflowed)
616 Warn(instrprof_error::counter_overflow);
617 ++I;
618 continue;
619 }
620 ValueData.insert(I, *J);
621 }
622 }
623
scale(uint64_t Weight,function_ref<void (instrprof_error)> Warn)624 void InstrProfValueSiteRecord::scale(uint64_t Weight,
625 function_ref<void(instrprof_error)> Warn) {
626 for (auto I = ValueData.begin(), IE = ValueData.end(); I != IE; ++I) {
627 bool Overflowed;
628 I->Count = SaturatingMultiply(I->Count, Weight, &Overflowed);
629 if (Overflowed)
630 Warn(instrprof_error::counter_overflow);
631 }
632 }
633
634 // Merge Value Profile data from Src record to this record for ValueKind.
635 // Scale merged value counts by \p Weight.
mergeValueProfData(uint32_t ValueKind,InstrProfRecord & Src,uint64_t Weight,function_ref<void (instrprof_error)> Warn)636 void InstrProfRecord::mergeValueProfData(
637 uint32_t ValueKind, InstrProfRecord &Src, uint64_t Weight,
638 function_ref<void(instrprof_error)> Warn) {
639 uint32_t ThisNumValueSites = getNumValueSites(ValueKind);
640 uint32_t OtherNumValueSites = Src.getNumValueSites(ValueKind);
641 if (ThisNumValueSites != OtherNumValueSites) {
642 Warn(instrprof_error::value_site_count_mismatch);
643 return;
644 }
645 if (!ThisNumValueSites)
646 return;
647 std::vector<InstrProfValueSiteRecord> &ThisSiteRecords =
648 getOrCreateValueSitesForKind(ValueKind);
649 MutableArrayRef<InstrProfValueSiteRecord> OtherSiteRecords =
650 Src.getValueSitesForKind(ValueKind);
651 for (uint32_t I = 0; I < ThisNumValueSites; I++)
652 ThisSiteRecords[I].merge(OtherSiteRecords[I], Weight, Warn);
653 }
654
merge(InstrProfRecord & Other,uint64_t Weight,function_ref<void (instrprof_error)> Warn)655 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight,
656 function_ref<void(instrprof_error)> Warn) {
657 // If the number of counters doesn't match we either have bad data
658 // or a hash collision.
659 if (Counts.size() != Other.Counts.size()) {
660 Warn(instrprof_error::count_mismatch);
661 return;
662 }
663
664 for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
665 bool Overflowed;
666 Counts[I] =
667 SaturatingMultiplyAdd(Other.Counts[I], Weight, Counts[I], &Overflowed);
668 if (Overflowed)
669 Warn(instrprof_error::counter_overflow);
670 }
671
672 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
673 mergeValueProfData(Kind, Other, Weight, Warn);
674 }
675
scaleValueProfData(uint32_t ValueKind,uint64_t Weight,function_ref<void (instrprof_error)> Warn)676 void InstrProfRecord::scaleValueProfData(
677 uint32_t ValueKind, uint64_t Weight,
678 function_ref<void(instrprof_error)> Warn) {
679 for (auto &R : getValueSitesForKind(ValueKind))
680 R.scale(Weight, Warn);
681 }
682
scale(uint64_t Weight,function_ref<void (instrprof_error)> Warn)683 void InstrProfRecord::scale(uint64_t Weight,
684 function_ref<void(instrprof_error)> Warn) {
685 for (auto &Count : this->Counts) {
686 bool Overflowed;
687 Count = SaturatingMultiply(Count, Weight, &Overflowed);
688 if (Overflowed)
689 Warn(instrprof_error::counter_overflow);
690 }
691 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
692 scaleValueProfData(Kind, Weight, Warn);
693 }
694
695 // Map indirect call target name hash to name string.
remapValue(uint64_t Value,uint32_t ValueKind,InstrProfSymtab * SymTab)696 uint64_t InstrProfRecord::remapValue(uint64_t Value, uint32_t ValueKind,
697 InstrProfSymtab *SymTab) {
698 if (!SymTab)
699 return Value;
700
701 if (ValueKind == IPVK_IndirectCallTarget)
702 return SymTab->getFunctionHashFromAddress(Value);
703
704 return Value;
705 }
706
addValueData(uint32_t ValueKind,uint32_t Site,InstrProfValueData * VData,uint32_t N,InstrProfSymtab * ValueMap)707 void InstrProfRecord::addValueData(uint32_t ValueKind, uint32_t Site,
708 InstrProfValueData *VData, uint32_t N,
709 InstrProfSymtab *ValueMap) {
710 for (uint32_t I = 0; I < N; I++) {
711 VData[I].Value = remapValue(VData[I].Value, ValueKind, ValueMap);
712 }
713 std::vector<InstrProfValueSiteRecord> &ValueSites =
714 getOrCreateValueSitesForKind(ValueKind);
715 if (N == 0)
716 ValueSites.emplace_back();
717 else
718 ValueSites.emplace_back(VData, VData + N);
719 }
720
721 #define INSTR_PROF_COMMON_API_IMPL
722 #include "llvm/ProfileData/InstrProfData.inc"
723
724 /*!
725 * ValueProfRecordClosure Interface implementation for InstrProfRecord
726 * class. These C wrappers are used as adaptors so that C++ code can be
727 * invoked as callbacks.
728 */
getNumValueKindsInstrProf(const void * Record)729 uint32_t getNumValueKindsInstrProf(const void *Record) {
730 return reinterpret_cast<const InstrProfRecord *>(Record)->getNumValueKinds();
731 }
732
getNumValueSitesInstrProf(const void * Record,uint32_t VKind)733 uint32_t getNumValueSitesInstrProf(const void *Record, uint32_t VKind) {
734 return reinterpret_cast<const InstrProfRecord *>(Record)
735 ->getNumValueSites(VKind);
736 }
737
getNumValueDataInstrProf(const void * Record,uint32_t VKind)738 uint32_t getNumValueDataInstrProf(const void *Record, uint32_t VKind) {
739 return reinterpret_cast<const InstrProfRecord *>(Record)
740 ->getNumValueData(VKind);
741 }
742
getNumValueDataForSiteInstrProf(const void * R,uint32_t VK,uint32_t S)743 uint32_t getNumValueDataForSiteInstrProf(const void *R, uint32_t VK,
744 uint32_t S) {
745 return reinterpret_cast<const InstrProfRecord *>(R)
746 ->getNumValueDataForSite(VK, S);
747 }
748
getValueForSiteInstrProf(const void * R,InstrProfValueData * Dst,uint32_t K,uint32_t S)749 void getValueForSiteInstrProf(const void *R, InstrProfValueData *Dst,
750 uint32_t K, uint32_t S) {
751 reinterpret_cast<const InstrProfRecord *>(R)->getValueForSite(Dst, K, S);
752 }
753
allocValueProfDataInstrProf(size_t TotalSizeInBytes)754 ValueProfData *allocValueProfDataInstrProf(size_t TotalSizeInBytes) {
755 ValueProfData *VD =
756 (ValueProfData *)(new (::operator new(TotalSizeInBytes)) ValueProfData());
757 memset(VD, 0, TotalSizeInBytes);
758 return VD;
759 }
760
761 static ValueProfRecordClosure InstrProfRecordClosure = {
762 nullptr,
763 getNumValueKindsInstrProf,
764 getNumValueSitesInstrProf,
765 getNumValueDataInstrProf,
766 getNumValueDataForSiteInstrProf,
767 nullptr,
768 getValueForSiteInstrProf,
769 allocValueProfDataInstrProf};
770
771 // Wrapper implementation using the closure mechanism.
getSize(const InstrProfRecord & Record)772 uint32_t ValueProfData::getSize(const InstrProfRecord &Record) {
773 auto Closure = InstrProfRecordClosure;
774 Closure.Record = &Record;
775 return getValueProfDataSize(&Closure);
776 }
777
778 // Wrapper implementation using the closure mechanism.
779 std::unique_ptr<ValueProfData>
serializeFrom(const InstrProfRecord & Record)780 ValueProfData::serializeFrom(const InstrProfRecord &Record) {
781 InstrProfRecordClosure.Record = &Record;
782
783 std::unique_ptr<ValueProfData> VPD(
784 serializeValueProfDataFrom(&InstrProfRecordClosure, nullptr));
785 return VPD;
786 }
787
deserializeTo(InstrProfRecord & Record,InstrProfSymtab * SymTab)788 void ValueProfRecord::deserializeTo(InstrProfRecord &Record,
789 InstrProfSymtab *SymTab) {
790 Record.reserveSites(Kind, NumValueSites);
791
792 InstrProfValueData *ValueData = getValueProfRecordValueData(this);
793 for (uint64_t VSite = 0; VSite < NumValueSites; ++VSite) {
794 uint8_t ValueDataCount = this->SiteCountArray[VSite];
795 Record.addValueData(Kind, VSite, ValueData, ValueDataCount, SymTab);
796 ValueData += ValueDataCount;
797 }
798 }
799
800 // For writing/serializing, Old is the host endianness, and New is
801 // byte order intended on disk. For Reading/deserialization, Old
802 // is the on-disk source endianness, and New is the host endianness.
swapBytes(support::endianness Old,support::endianness New)803 void ValueProfRecord::swapBytes(support::endianness Old,
804 support::endianness New) {
805 using namespace support;
806
807 if (Old == New)
808 return;
809
810 if (getHostEndianness() != Old) {
811 sys::swapByteOrder<uint32_t>(NumValueSites);
812 sys::swapByteOrder<uint32_t>(Kind);
813 }
814 uint32_t ND = getValueProfRecordNumValueData(this);
815 InstrProfValueData *VD = getValueProfRecordValueData(this);
816
817 // No need to swap byte array: SiteCountArrray.
818 for (uint32_t I = 0; I < ND; I++) {
819 sys::swapByteOrder<uint64_t>(VD[I].Value);
820 sys::swapByteOrder<uint64_t>(VD[I].Count);
821 }
822 if (getHostEndianness() == Old) {
823 sys::swapByteOrder<uint32_t>(NumValueSites);
824 sys::swapByteOrder<uint32_t>(Kind);
825 }
826 }
827
deserializeTo(InstrProfRecord & Record,InstrProfSymtab * SymTab)828 void ValueProfData::deserializeTo(InstrProfRecord &Record,
829 InstrProfSymtab *SymTab) {
830 if (NumValueKinds == 0)
831 return;
832
833 ValueProfRecord *VR = getFirstValueProfRecord(this);
834 for (uint32_t K = 0; K < NumValueKinds; K++) {
835 VR->deserializeTo(Record, SymTab);
836 VR = getValueProfRecordNext(VR);
837 }
838 }
839
840 template <class T>
swapToHostOrder(const unsigned char * & D,support::endianness Orig)841 static T swapToHostOrder(const unsigned char *&D, support::endianness Orig) {
842 using namespace support;
843
844 if (Orig == little)
845 return endian::readNext<T, little, unaligned>(D);
846 else
847 return endian::readNext<T, big, unaligned>(D);
848 }
849
allocValueProfData(uint32_t TotalSize)850 static std::unique_ptr<ValueProfData> allocValueProfData(uint32_t TotalSize) {
851 return std::unique_ptr<ValueProfData>(new (::operator new(TotalSize))
852 ValueProfData());
853 }
854
checkIntegrity()855 Error ValueProfData::checkIntegrity() {
856 if (NumValueKinds > IPVK_Last + 1)
857 return make_error<InstrProfError>(instrprof_error::malformed);
858 // Total size needs to be mulltiple of quadword size.
859 if (TotalSize % sizeof(uint64_t))
860 return make_error<InstrProfError>(instrprof_error::malformed);
861
862 ValueProfRecord *VR = getFirstValueProfRecord(this);
863 for (uint32_t K = 0; K < this->NumValueKinds; K++) {
864 if (VR->Kind > IPVK_Last)
865 return make_error<InstrProfError>(instrprof_error::malformed);
866 VR = getValueProfRecordNext(VR);
867 if ((char *)VR - (char *)this > (ptrdiff_t)TotalSize)
868 return make_error<InstrProfError>(instrprof_error::malformed);
869 }
870 return Error::success();
871 }
872
873 Expected<std::unique_ptr<ValueProfData>>
getValueProfData(const unsigned char * D,const unsigned char * const BufferEnd,support::endianness Endianness)874 ValueProfData::getValueProfData(const unsigned char *D,
875 const unsigned char *const BufferEnd,
876 support::endianness Endianness) {
877 using namespace support;
878
879 if (D + sizeof(ValueProfData) > BufferEnd)
880 return make_error<InstrProfError>(instrprof_error::truncated);
881
882 const unsigned char *Header = D;
883 uint32_t TotalSize = swapToHostOrder<uint32_t>(Header, Endianness);
884 if (D + TotalSize > BufferEnd)
885 return make_error<InstrProfError>(instrprof_error::too_large);
886
887 std::unique_ptr<ValueProfData> VPD = allocValueProfData(TotalSize);
888 memcpy(VPD.get(), D, TotalSize);
889 // Byte swap.
890 VPD->swapBytesToHost(Endianness);
891
892 Error E = VPD->checkIntegrity();
893 if (E)
894 return std::move(E);
895
896 return std::move(VPD);
897 }
898
swapBytesToHost(support::endianness Endianness)899 void ValueProfData::swapBytesToHost(support::endianness Endianness) {
900 using namespace support;
901
902 if (Endianness == getHostEndianness())
903 return;
904
905 sys::swapByteOrder<uint32_t>(TotalSize);
906 sys::swapByteOrder<uint32_t>(NumValueKinds);
907
908 ValueProfRecord *VR = getFirstValueProfRecord(this);
909 for (uint32_t K = 0; K < NumValueKinds; K++) {
910 VR->swapBytes(Endianness, getHostEndianness());
911 VR = getValueProfRecordNext(VR);
912 }
913 }
914
swapBytesFromHost(support::endianness Endianness)915 void ValueProfData::swapBytesFromHost(support::endianness Endianness) {
916 using namespace support;
917
918 if (Endianness == getHostEndianness())
919 return;
920
921 ValueProfRecord *VR = getFirstValueProfRecord(this);
922 for (uint32_t K = 0; K < NumValueKinds; K++) {
923 ValueProfRecord *NVR = getValueProfRecordNext(VR);
924 VR->swapBytes(getHostEndianness(), Endianness);
925 VR = NVR;
926 }
927 sys::swapByteOrder<uint32_t>(TotalSize);
928 sys::swapByteOrder<uint32_t>(NumValueKinds);
929 }
930
annotateValueSite(Module & M,Instruction & Inst,const InstrProfRecord & InstrProfR,InstrProfValueKind ValueKind,uint32_t SiteIdx,uint32_t MaxMDCount)931 void annotateValueSite(Module &M, Instruction &Inst,
932 const InstrProfRecord &InstrProfR,
933 InstrProfValueKind ValueKind, uint32_t SiteIdx,
934 uint32_t MaxMDCount) {
935 uint32_t NV = InstrProfR.getNumValueDataForSite(ValueKind, SiteIdx);
936 if (!NV)
937 return;
938
939 uint64_t Sum = 0;
940 std::unique_ptr<InstrProfValueData[]> VD =
941 InstrProfR.getValueForSite(ValueKind, SiteIdx, &Sum);
942
943 ArrayRef<InstrProfValueData> VDs(VD.get(), NV);
944 annotateValueSite(M, Inst, VDs, Sum, ValueKind, MaxMDCount);
945 }
946
annotateValueSite(Module & M,Instruction & Inst,ArrayRef<InstrProfValueData> VDs,uint64_t Sum,InstrProfValueKind ValueKind,uint32_t MaxMDCount)947 void annotateValueSite(Module &M, Instruction &Inst,
948 ArrayRef<InstrProfValueData> VDs,
949 uint64_t Sum, InstrProfValueKind ValueKind,
950 uint32_t MaxMDCount) {
951 LLVMContext &Ctx = M.getContext();
952 MDBuilder MDHelper(Ctx);
953 SmallVector<Metadata *, 3> Vals;
954 // Tag
955 Vals.push_back(MDHelper.createString("VP"));
956 // Value Kind
957 Vals.push_back(MDHelper.createConstant(
958 ConstantInt::get(Type::getInt32Ty(Ctx), ValueKind)));
959 // Total Count
960 Vals.push_back(
961 MDHelper.createConstant(ConstantInt::get(Type::getInt64Ty(Ctx), Sum)));
962
963 // Value Profile Data
964 uint32_t MDCount = MaxMDCount;
965 for (auto &VD : VDs) {
966 Vals.push_back(MDHelper.createConstant(
967 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Value)));
968 Vals.push_back(MDHelper.createConstant(
969 ConstantInt::get(Type::getInt64Ty(Ctx), VD.Count)));
970 if (--MDCount == 0)
971 break;
972 }
973 Inst.setMetadata(LLVMContext::MD_prof, MDNode::get(Ctx, Vals));
974 }
975
getValueProfDataFromInst(const Instruction & Inst,InstrProfValueKind ValueKind,uint32_t MaxNumValueData,InstrProfValueData ValueData[],uint32_t & ActualNumValueData,uint64_t & TotalC)976 bool getValueProfDataFromInst(const Instruction &Inst,
977 InstrProfValueKind ValueKind,
978 uint32_t MaxNumValueData,
979 InstrProfValueData ValueData[],
980 uint32_t &ActualNumValueData, uint64_t &TotalC) {
981 MDNode *MD = Inst.getMetadata(LLVMContext::MD_prof);
982 if (!MD)
983 return false;
984
985 unsigned NOps = MD->getNumOperands();
986
987 if (NOps < 5)
988 return false;
989
990 // Operand 0 is a string tag "VP":
991 MDString *Tag = cast<MDString>(MD->getOperand(0));
992 if (!Tag)
993 return false;
994
995 if (!Tag->getString().equals("VP"))
996 return false;
997
998 // Now check kind:
999 ConstantInt *KindInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
1000 if (!KindInt)
1001 return false;
1002 if (KindInt->getZExtValue() != ValueKind)
1003 return false;
1004
1005 // Get total count
1006 ConstantInt *TotalCInt = mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
1007 if (!TotalCInt)
1008 return false;
1009 TotalC = TotalCInt->getZExtValue();
1010
1011 ActualNumValueData = 0;
1012
1013 for (unsigned I = 3; I < NOps; I += 2) {
1014 if (ActualNumValueData >= MaxNumValueData)
1015 break;
1016 ConstantInt *Value = mdconst::dyn_extract<ConstantInt>(MD->getOperand(I));
1017 ConstantInt *Count =
1018 mdconst::dyn_extract<ConstantInt>(MD->getOperand(I + 1));
1019 if (!Value || !Count)
1020 return false;
1021 ValueData[ActualNumValueData].Value = Value->getZExtValue();
1022 ValueData[ActualNumValueData].Count = Count->getZExtValue();
1023 ActualNumValueData++;
1024 }
1025 return true;
1026 }
1027
getPGOFuncNameMetadata(const Function & F)1028 MDNode *getPGOFuncNameMetadata(const Function &F) {
1029 return F.getMetadata(getPGOFuncNameMetadataName());
1030 }
1031
createPGOFuncNameMetadata(Function & F,StringRef PGOFuncName)1032 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName) {
1033 // Only for internal linkage functions.
1034 if (PGOFuncName == F.getName())
1035 return;
1036 // Don't create duplicated meta-data.
1037 if (getPGOFuncNameMetadata(F))
1038 return;
1039 LLVMContext &C = F.getContext();
1040 MDNode *N = MDNode::get(C, MDString::get(C, PGOFuncName));
1041 F.setMetadata(getPGOFuncNameMetadataName(), N);
1042 }
1043
needsComdatForCounter(const Function & F,const Module & M)1044 bool needsComdatForCounter(const Function &F, const Module &M) {
1045 if (F.hasComdat())
1046 return true;
1047
1048 if (!Triple(M.getTargetTriple()).supportsCOMDAT())
1049 return false;
1050
1051 // See createPGOFuncNameVar for more details. To avoid link errors, profile
1052 // counters for function with available_externally linkage needs to be changed
1053 // to linkonce linkage. On ELF based systems, this leads to weak symbols to be
1054 // created. Without using comdat, duplicate entries won't be removed by the
1055 // linker leading to increased data segement size and raw profile size. Even
1056 // worse, since the referenced counter from profile per-function data object
1057 // will be resolved to the common strong definition, the profile counts for
1058 // available_externally functions will end up being duplicated in raw profile
1059 // data. This can result in distorted profile as the counts of those dups
1060 // will be accumulated by the profile merger.
1061 GlobalValue::LinkageTypes Linkage = F.getLinkage();
1062 if (Linkage != GlobalValue::ExternalWeakLinkage &&
1063 Linkage != GlobalValue::AvailableExternallyLinkage)
1064 return false;
1065
1066 return true;
1067 }
1068
1069 // Check if INSTR_PROF_RAW_VERSION_VAR is defined.
isIRPGOFlagSet(const Module * M)1070 bool isIRPGOFlagSet(const Module *M) {
1071 auto IRInstrVar =
1072 M->getNamedGlobal(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
1073 if (!IRInstrVar || IRInstrVar->isDeclaration() ||
1074 IRInstrVar->hasLocalLinkage())
1075 return false;
1076
1077 // Check if the flag is set.
1078 if (!IRInstrVar->hasInitializer())
1079 return false;
1080
1081 auto *InitVal = dyn_cast_or_null<ConstantInt>(IRInstrVar->getInitializer());
1082 if (!InitVal)
1083 return false;
1084 return (InitVal->getZExtValue() & VARIANT_MASK_IR_PROF) != 0;
1085 }
1086
1087 // Check if we can safely rename this Comdat function.
canRenameComdatFunc(const Function & F,bool CheckAddressTaken)1088 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
1089 if (F.getName().empty())
1090 return false;
1091 if (!needsComdatForCounter(F, *(F.getParent())))
1092 return false;
1093 // Unsafe to rename the address-taken function (which can be used in
1094 // function comparison).
1095 if (CheckAddressTaken && F.hasAddressTaken())
1096 return false;
1097 // Only safe to do if this function may be discarded if it is not used
1098 // in the compilation unit.
1099 if (!GlobalValue::isDiscardableIfUnused(F.getLinkage()))
1100 return false;
1101
1102 // For AvailableExternallyLinkage functions.
1103 if (!F.hasComdat()) {
1104 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
1105 return true;
1106 }
1107 return true;
1108 }
1109
1110 // Parse the value profile options.
getMemOPSizeRangeFromOption(StringRef MemOPSizeRange,int64_t & RangeStart,int64_t & RangeLast)1111 void getMemOPSizeRangeFromOption(StringRef MemOPSizeRange, int64_t &RangeStart,
1112 int64_t &RangeLast) {
1113 static const int64_t DefaultMemOPSizeRangeStart = 0;
1114 static const int64_t DefaultMemOPSizeRangeLast = 8;
1115 RangeStart = DefaultMemOPSizeRangeStart;
1116 RangeLast = DefaultMemOPSizeRangeLast;
1117
1118 if (!MemOPSizeRange.empty()) {
1119 auto Pos = MemOPSizeRange.find(':');
1120 if (Pos != std::string::npos) {
1121 if (Pos > 0)
1122 MemOPSizeRange.substr(0, Pos).getAsInteger(10, RangeStart);
1123 if (Pos < MemOPSizeRange.size() - 1)
1124 MemOPSizeRange.substr(Pos + 1).getAsInteger(10, RangeLast);
1125 } else
1126 MemOPSizeRange.getAsInteger(10, RangeLast);
1127 }
1128 assert(RangeLast >= RangeStart);
1129 }
1130
1131 // Create a COMDAT variable INSTR_PROF_RAW_VERSION_VAR to make the runtime
1132 // aware this is an ir_level profile so it can set the version flag.
createIRLevelProfileFlagVar(Module & M,bool IsCS)1133 void createIRLevelProfileFlagVar(Module &M, bool IsCS) {
1134 const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
1135 Type *IntTy64 = Type::getInt64Ty(M.getContext());
1136 uint64_t ProfileVersion = (INSTR_PROF_RAW_VERSION | VARIANT_MASK_IR_PROF);
1137 if (IsCS)
1138 ProfileVersion |= VARIANT_MASK_CSIR_PROF;
1139 auto IRLevelVersionVariable = new GlobalVariable(
1140 M, IntTy64, true, GlobalValue::WeakAnyLinkage,
1141 Constant::getIntegerValue(IntTy64, APInt(64, ProfileVersion)), VarName);
1142 IRLevelVersionVariable->setVisibility(GlobalValue::DefaultVisibility);
1143 Triple TT(M.getTargetTriple());
1144 if (TT.supportsCOMDAT()) {
1145 IRLevelVersionVariable->setLinkage(GlobalValue::ExternalLinkage);
1146 IRLevelVersionVariable->setComdat(M.getOrInsertComdat(VarName));
1147 }
1148 }
1149
1150 // Create the variable for the profile file name.
createProfileFileNameVar(Module & M,StringRef InstrProfileOutput)1151 void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput) {
1152 if (InstrProfileOutput.empty())
1153 return;
1154 Constant *ProfileNameConst =
1155 ConstantDataArray::getString(M.getContext(), InstrProfileOutput, true);
1156 GlobalVariable *ProfileNameVar = new GlobalVariable(
1157 M, ProfileNameConst->getType(), true, GlobalValue::WeakAnyLinkage,
1158 ProfileNameConst, INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR));
1159 Triple TT(M.getTargetTriple());
1160 if (TT.supportsCOMDAT()) {
1161 ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage);
1162 ProfileNameVar->setComdat(M.getOrInsertComdat(
1163 StringRef(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_NAME_VAR))));
1164 }
1165 }
1166
accumulateCounts(const std::string & BaseFilename,const std::string & TestFilename,bool IsCS)1167 Error OverlapStats::accumulateCounts(const std::string &BaseFilename,
1168 const std::string &TestFilename,
1169 bool IsCS) {
1170 auto getProfileSum = [IsCS](const std::string &Filename,
1171 CountSumOrPercent &Sum) -> Error {
1172 auto ReaderOrErr = InstrProfReader::create(Filename);
1173 if (Error E = ReaderOrErr.takeError()) {
1174 return E;
1175 }
1176 auto Reader = std::move(ReaderOrErr.get());
1177 Reader->accumulateCounts(Sum, IsCS);
1178 return Error::success();
1179 };
1180 auto Ret = getProfileSum(BaseFilename, Base);
1181 if (Ret)
1182 return Ret;
1183 Ret = getProfileSum(TestFilename, Test);
1184 if (Ret)
1185 return Ret;
1186 this->BaseFilename = &BaseFilename;
1187 this->TestFilename = &TestFilename;
1188 Valid = true;
1189 return Error::success();
1190 }
1191
addOneMismatch(const CountSumOrPercent & MismatchFunc)1192 void OverlapStats::addOneMismatch(const CountSumOrPercent &MismatchFunc) {
1193 Mismatch.NumEntries += 1;
1194 Mismatch.CountSum += MismatchFunc.CountSum / Test.CountSum;
1195 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1196 if (Test.ValueCounts[I] >= 1.0f)
1197 Mismatch.ValueCounts[I] +=
1198 MismatchFunc.ValueCounts[I] / Test.ValueCounts[I];
1199 }
1200 }
1201
addOneUnique(const CountSumOrPercent & UniqueFunc)1202 void OverlapStats::addOneUnique(const CountSumOrPercent &UniqueFunc) {
1203 Unique.NumEntries += 1;
1204 Unique.CountSum += UniqueFunc.CountSum / Test.CountSum;
1205 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1206 if (Test.ValueCounts[I] >= 1.0f)
1207 Unique.ValueCounts[I] += UniqueFunc.ValueCounts[I] / Test.ValueCounts[I];
1208 }
1209 }
1210
dump(raw_fd_ostream & OS) const1211 void OverlapStats::dump(raw_fd_ostream &OS) const {
1212 if (!Valid)
1213 return;
1214
1215 const char *EntryName =
1216 (Level == ProgramLevel ? "functions" : "edge counters");
1217 if (Level == ProgramLevel) {
1218 OS << "Profile overlap infomation for base_profile: " << *BaseFilename
1219 << " and test_profile: " << *TestFilename << "\nProgram level:\n";
1220 } else {
1221 OS << "Function level:\n"
1222 << " Function: " << FuncName << " (Hash=" << FuncHash << ")\n";
1223 }
1224
1225 OS << " # of " << EntryName << " overlap: " << Overlap.NumEntries << "\n";
1226 if (Mismatch.NumEntries)
1227 OS << " # of " << EntryName << " mismatch: " << Mismatch.NumEntries
1228 << "\n";
1229 if (Unique.NumEntries)
1230 OS << " # of " << EntryName
1231 << " only in test_profile: " << Unique.NumEntries << "\n";
1232
1233 OS << " Edge profile overlap: " << format("%.3f%%", Overlap.CountSum * 100)
1234 << "\n";
1235 if (Mismatch.NumEntries)
1236 OS << " Mismatched count percentage (Edge): "
1237 << format("%.3f%%", Mismatch.CountSum * 100) << "\n";
1238 if (Unique.NumEntries)
1239 OS << " Percentage of Edge profile only in test_profile: "
1240 << format("%.3f%%", Unique.CountSum * 100) << "\n";
1241 OS << " Edge profile base count sum: " << format("%.0f", Base.CountSum)
1242 << "\n"
1243 << " Edge profile test count sum: " << format("%.0f", Test.CountSum)
1244 << "\n";
1245
1246 for (unsigned I = 0; I < IPVK_Last - IPVK_First + 1; I++) {
1247 if (Base.ValueCounts[I] < 1.0f && Test.ValueCounts[I] < 1.0f)
1248 continue;
1249 char ProfileKindName[20];
1250 switch (I) {
1251 case IPVK_IndirectCallTarget:
1252 strncpy(ProfileKindName, "IndirectCall", 19);
1253 break;
1254 case IPVK_MemOPSize:
1255 strncpy(ProfileKindName, "MemOP", 19);
1256 break;
1257 default:
1258 snprintf(ProfileKindName, 19, "VP[%d]", I);
1259 break;
1260 }
1261 OS << " " << ProfileKindName
1262 << " profile overlap: " << format("%.3f%%", Overlap.ValueCounts[I] * 100)
1263 << "\n";
1264 if (Mismatch.NumEntries)
1265 OS << " Mismatched count percentage (" << ProfileKindName
1266 << "): " << format("%.3f%%", Mismatch.ValueCounts[I] * 100) << "\n";
1267 if (Unique.NumEntries)
1268 OS << " Percentage of " << ProfileKindName
1269 << " profile only in test_profile: "
1270 << format("%.3f%%", Unique.ValueCounts[I] * 100) << "\n";
1271 OS << " " << ProfileKindName
1272 << " profile base count sum: " << format("%.0f", Base.ValueCounts[I])
1273 << "\n"
1274 << " " << ProfileKindName
1275 << " profile test count sum: " << format("%.0f", Test.ValueCounts[I])
1276 << "\n";
1277 }
1278 }
1279
1280 } // end namespace llvm
1281