1 //===- GCOVProfiling.cpp - Insert edge counters for gcov profiling --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements GCOV-style profiling. When this pass is run it emits
11 // "gcno" files next to the existing source, and instruments the code that runs
12 // to records the edges between blocks that run and emit a complementary "gcda"
15 //===----------------------------------------------------------------------===//
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/Hashing.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/Sequence.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/ADT/StringExtras.h"
23 #include "llvm/ADT/StringMap.h"
24 #include "llvm/Analysis/EHPersonalities.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/IR/CFG.h"
27 #include "llvm/IR/DebugInfo.h"
28 #include "llvm/IR/DebugLoc.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/FileSystem.h"
38 #include "llvm/Support/Path.h"
39 #include "llvm/Support/Regex.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Instrumentation.h"
42 #include "llvm/Transforms/Instrumentation/GCOVProfiler.h"
43 #include "llvm/Transforms/Utils/ModuleUtils.h"
50 #define DEBUG_TYPE "insert-gcov-profiling"
52 static cl::opt<std::string>
53 DefaultGCOVVersion("default-gcov-version", cl::init("402*"), cl::Hidden,
55 static cl::opt<bool> DefaultExitBlockBeforeBody("gcov-exit-block-before-body",
56 cl::init(false), cl::Hidden);
58 GCOVOptions GCOVOptions::getDefault() {
60 Options.EmitNotes = true;
61 Options.EmitData = true;
62 Options.UseCfgChecksum = false;
63 Options.NoRedZone = false;
64 Options.FunctionNamesInData = true;
65 Options.ExitBlockBeforeBody = DefaultExitBlockBeforeBody;
67 if (DefaultGCOVVersion.size() != 4) {
68 llvm::report_fatal_error(std::string("Invalid -default-gcov-version: ") +
71 memcpy(Options.Version, DefaultGCOVVersion.c_str(), 4);
80 GCOVProfiler() : GCOVProfiler(GCOVOptions::getDefault()) {}
81 GCOVProfiler(const GCOVOptions &Opts) : Options(Opts) {
82 assert((Options.EmitNotes || Options.EmitData) &&
83 "GCOVProfiler asked to do nothing?");
84 ReversedVersion[0] = Options.Version[3];
85 ReversedVersion[1] = Options.Version[2];
86 ReversedVersion[2] = Options.Version[1];
87 ReversedVersion[3] = Options.Version[0];
88 ReversedVersion[4] = '\0';
90 bool runOnModule(Module &M, const TargetLibraryInfo &TLI);
93 // Create the .gcno files for the Module based on DebugInfo.
94 void emitProfileNotes();
96 // Modify the program to track transitions along edges and call into the
97 // profiling runtime to emit .gcda files when run.
98 bool emitProfileArcs();
100 bool isFunctionInstrumented(const Function &F);
101 std::vector<Regex> createRegexesFromString(StringRef RegexesStr);
102 static bool doesFilenameMatchARegex(StringRef Filename,
103 std::vector<Regex> &Regexes);
105 // Get pointers to the functions in the runtime library.
106 Constant *getStartFileFunc();
107 Constant *getEmitFunctionFunc();
108 Constant *getEmitArcsFunc();
109 Constant *getSummaryInfoFunc();
110 Constant *getEndFileFunc();
112 // Add the function to write out all our counters to the global destructor
115 insertCounterWriteout(ArrayRef<std::pair<GlobalVariable *, MDNode *>>);
116 Function *insertFlush(ArrayRef<std::pair<GlobalVariable *, MDNode *>>);
118 void AddFlushBeforeForkAndExec();
120 enum class GCovFileType { GCNO, GCDA };
121 std::string mangleName(const DICompileUnit *CU, GCovFileType FileType);
125 // Reversed, NUL-terminated copy of Options.Version.
126 char ReversedVersion[5];
127 // Checksum, produced by hash of EdgeDestinations
128 SmallVector<uint32_t, 4> FileChecksums;
131 const TargetLibraryInfo *TLI;
133 SmallVector<std::unique_ptr<GCOVFunction>, 16> Funcs;
134 std::vector<Regex> FilterRe;
135 std::vector<Regex> ExcludeRe;
136 StringMap<bool> InstrumentedFiles;
139 class GCOVProfilerLegacyPass : public ModulePass {
142 GCOVProfilerLegacyPass()
143 : GCOVProfilerLegacyPass(GCOVOptions::getDefault()) {}
144 GCOVProfilerLegacyPass(const GCOVOptions &Opts)
145 : ModulePass(ID), Profiler(Opts) {
146 initializeGCOVProfilerLegacyPassPass(*PassRegistry::getPassRegistry());
148 StringRef getPassName() const override { return "GCOV Profiler"; }
150 bool runOnModule(Module &M) override {
151 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
152 return Profiler.runOnModule(M, TLI);
155 void getAnalysisUsage(AnalysisUsage &AU) const override {
156 AU.addRequired<TargetLibraryInfoWrapperPass>();
160 GCOVProfiler Profiler;
164 char GCOVProfilerLegacyPass::ID = 0;
165 INITIALIZE_PASS_BEGIN(
166 GCOVProfilerLegacyPass, "insert-gcov-profiling",
167 "Insert instrumentation for GCOV profiling", false, false)
168 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
170 GCOVProfilerLegacyPass, "insert-gcov-profiling",
171 "Insert instrumentation for GCOV profiling", false, false)
173 ModulePass *llvm::createGCOVProfilerPass(const GCOVOptions &Options) {
174 return new GCOVProfilerLegacyPass(Options);
177 static StringRef getFunctionName(const DISubprogram *SP) {
178 if (!SP->getLinkageName().empty())
179 return SP->getLinkageName();
180 return SP->getName();
183 /// Extract a filename for a DISubprogram.
185 /// Prefer relative paths in the coverage notes. Clang also may split
186 /// up absolute paths into a directory and filename component. When
187 /// the relative path doesn't exist, reconstruct the absolute path.
188 static SmallString<128> getFilename(const DISubprogram *SP) {
189 SmallString<128> Path;
190 StringRef RelPath = SP->getFilename();
191 if (sys::fs::exists(RelPath))
194 sys::path::append(Path, SP->getDirectory(), SP->getFilename());
201 static const char *const LinesTag;
202 static const char *const FunctionTag;
203 static const char *const BlockTag;
204 static const char *const EdgeTag;
206 GCOVRecord() = default;
208 void writeBytes(const char *Bytes, int Size) {
209 os->write(Bytes, Size);
212 void write(uint32_t i) {
213 writeBytes(reinterpret_cast<char*>(&i), 4);
216 // Returns the length measured in 4-byte blocks that will be used to
217 // represent this string in a GCOV file
218 static unsigned lengthOfGCOVString(StringRef s) {
219 // A GCOV string is a length, followed by a NUL, then between 0 and 3 NULs
220 // padding out to the next 4-byte word. The length is measured in 4-byte
221 // words including padding, not bytes of actual string.
222 return (s.size() / 4) + 1;
225 void writeGCOVString(StringRef s) {
226 uint32_t Len = lengthOfGCOVString(s);
228 writeBytes(s.data(), s.size());
230 // Write 1 to 4 bytes of NUL padding.
231 assert((unsigned)(4 - (s.size() % 4)) > 0);
232 assert((unsigned)(4 - (s.size() % 4)) <= 4);
233 writeBytes("\0\0\0\0", 4 - (s.size() % 4));
238 const char *const GCOVRecord::LinesTag = "\0\0\x45\x01";
239 const char *const GCOVRecord::FunctionTag = "\0\0\0\1";
240 const char *const GCOVRecord::BlockTag = "\0\0\x41\x01";
241 const char *const GCOVRecord::EdgeTag = "\0\0\x43\x01";
246 // Constructed only by requesting it from a GCOVBlock, this object stores a
247 // list of line numbers and a single filename, representing lines that belong
249 class GCOVLines : public GCOVRecord {
251 void addLine(uint32_t Line) {
252 assert(Line != 0 && "Line zero is not a valid real line number.");
253 Lines.push_back(Line);
256 uint32_t length() const {
257 // Here 2 = 1 for string length + 1 for '0' id#.
258 return lengthOfGCOVString(Filename) + 2 + Lines.size();
263 writeGCOVString(Filename);
264 for (int i = 0, e = Lines.size(); i != e; ++i)
268 GCOVLines(StringRef F, raw_ostream *os)
274 std::string Filename;
275 SmallVector<uint32_t, 32> Lines;
279 // Represent a basic block in GCOV. Each block has a unique number in the
280 // function, number of lines belonging to each block, and a set of edges to
282 class GCOVBlock : public GCOVRecord {
284 GCOVLines &getFile(StringRef Filename) {
285 return LinesByFile.try_emplace(Filename, Filename, os).first->second;
288 void addEdge(GCOVBlock &Successor) {
289 OutEdges.push_back(&Successor);
294 SmallVector<StringMapEntry<GCOVLines> *, 32> SortedLinesByFile;
295 for (auto &I : LinesByFile) {
296 Len += I.second.length();
297 SortedLinesByFile.push_back(&I);
300 writeBytes(LinesTag, 4);
304 llvm::sort(SortedLinesByFile, [](StringMapEntry<GCOVLines> *LHS,
305 StringMapEntry<GCOVLines> *RHS) {
306 return LHS->getKey() < RHS->getKey();
308 for (auto &I : SortedLinesByFile)
309 I->getValue().writeOut();
314 GCOVBlock(const GCOVBlock &RHS) : GCOVRecord(RHS), Number(RHS.Number) {
315 // Only allow copy before edges and lines have been added. After that,
316 // there are inter-block pointers (eg: edges) that won't take kindly to
317 // blocks being copied or moved around.
318 assert(LinesByFile.empty());
319 assert(OutEdges.empty());
323 friend class GCOVFunction;
325 GCOVBlock(uint32_t Number, raw_ostream *os)
331 StringMap<GCOVLines> LinesByFile;
332 SmallVector<GCOVBlock *, 4> OutEdges;
335 // A function has a unique identifier, a checksum (we leave as zero) and a
336 // set of blocks and a map of edges between blocks. This is the only GCOV
337 // object users can construct, the blocks and lines will be rooted here.
338 class GCOVFunction : public GCOVRecord {
340 GCOVFunction(const DISubprogram *SP, Function *F, raw_ostream *os,
341 uint32_t Ident, bool UseCfgChecksum, bool ExitBlockBeforeBody)
342 : SP(SP), Ident(Ident), UseCfgChecksum(UseCfgChecksum), CfgChecksum(0),
346 LLVM_DEBUG(dbgs() << "Function: " << getFunctionName(SP) << "\n");
349 for (auto &BB : *F) {
350 // Skip index 1 if it's assigned to the ReturnBlock.
351 if (i == 1 && ExitBlockBeforeBody)
353 Blocks.insert(std::make_pair(&BB, GCOVBlock(i++, os)));
355 if (!ExitBlockBeforeBody)
356 ReturnBlock.Number = i;
358 std::string FunctionNameAndLine;
359 raw_string_ostream FNLOS(FunctionNameAndLine);
360 FNLOS << getFunctionName(SP) << SP->getLine();
362 FuncChecksum = hash_value(FunctionNameAndLine);
365 GCOVBlock &getBlock(BasicBlock *BB) {
366 return Blocks.find(BB)->second;
369 GCOVBlock &getReturnBlock() {
373 std::string getEdgeDestinations() {
374 std::string EdgeDestinations;
375 raw_string_ostream EDOS(EdgeDestinations);
376 Function *F = Blocks.begin()->first->getParent();
377 for (BasicBlock &I : *F) {
378 GCOVBlock &Block = getBlock(&I);
379 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i)
380 EDOS << Block.OutEdges[i]->Number;
382 return EdgeDestinations;
385 uint32_t getFuncChecksum() {
389 void setCfgChecksum(uint32_t Checksum) {
390 CfgChecksum = Checksum;
394 writeBytes(FunctionTag, 4);
395 SmallString<128> Filename = getFilename(SP);
396 uint32_t BlockLen = 1 + 1 + 1 + lengthOfGCOVString(getFunctionName(SP)) +
397 1 + lengthOfGCOVString(Filename) + 1;
405 writeGCOVString(getFunctionName(SP));
406 writeGCOVString(Filename);
407 write(SP->getLine());
409 // Emit count of blocks.
410 writeBytes(BlockTag, 4);
411 write(Blocks.size() + 1);
412 for (int i = 0, e = Blocks.size() + 1; i != e; ++i) {
413 write(0); // No flags on our blocks.
415 LLVM_DEBUG(dbgs() << Blocks.size() << " blocks.\n");
417 // Emit edges between blocks.
418 if (Blocks.empty()) return;
419 Function *F = Blocks.begin()->first->getParent();
420 for (BasicBlock &I : *F) {
421 GCOVBlock &Block = getBlock(&I);
422 if (Block.OutEdges.empty()) continue;
424 writeBytes(EdgeTag, 4);
425 write(Block.OutEdges.size() * 2 + 1);
427 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i) {
428 LLVM_DEBUG(dbgs() << Block.Number << " -> "
429 << Block.OutEdges[i]->Number << "\n");
430 write(Block.OutEdges[i]->Number);
431 write(0); // no flags
435 // Emit lines for each block.
436 for (BasicBlock &I : *F)
437 getBlock(&I).writeOut();
441 const DISubprogram *SP;
443 uint32_t FuncChecksum;
445 uint32_t CfgChecksum;
446 DenseMap<BasicBlock *, GCOVBlock> Blocks;
447 GCOVBlock ReturnBlock;
451 // RegexesStr is a string containing differents regex separated by a semi-colon.
452 // For example "foo\..*$;bar\..*$".
453 std::vector<Regex> GCOVProfiler::createRegexesFromString(StringRef RegexesStr) {
454 std::vector<Regex> Regexes;
455 while (!RegexesStr.empty()) {
456 std::pair<StringRef, StringRef> HeadTail = RegexesStr.split(';');
457 if (!HeadTail.first.empty()) {
458 Regex Re(HeadTail.first);
460 if (!Re.isValid(Err)) {
461 Ctx->emitError(Twine("Regex ") + HeadTail.first +
462 " is not valid: " + Err);
464 Regexes.emplace_back(std::move(Re));
466 RegexesStr = HeadTail.second;
471 bool GCOVProfiler::doesFilenameMatchARegex(StringRef Filename,
472 std::vector<Regex> &Regexes) {
473 for (Regex &Re : Regexes) {
474 if (Re.match(Filename)) {
481 bool GCOVProfiler::isFunctionInstrumented(const Function &F) {
482 if (FilterRe.empty() && ExcludeRe.empty()) {
485 SmallString<128> Filename = getFilename(F.getSubprogram());
486 auto It = InstrumentedFiles.find(Filename);
487 if (It != InstrumentedFiles.end()) {
491 SmallString<256> RealPath;
492 StringRef RealFilename;
495 // /usr/lib/gcc/x86_64-linux-gnu/8/../../../../include/c++/8/bits/*.h so for
496 // such a case we must get the real_path.
497 if (sys::fs::real_path(Filename, RealPath)) {
498 // real_path can fail with path like "foo.c".
499 RealFilename = Filename;
501 RealFilename = RealPath;
504 bool ShouldInstrument;
505 if (FilterRe.empty()) {
506 ShouldInstrument = !doesFilenameMatchARegex(RealFilename, ExcludeRe);
507 } else if (ExcludeRe.empty()) {
508 ShouldInstrument = doesFilenameMatchARegex(RealFilename, FilterRe);
510 ShouldInstrument = doesFilenameMatchARegex(RealFilename, FilterRe) &&
511 !doesFilenameMatchARegex(RealFilename, ExcludeRe);
513 InstrumentedFiles[Filename] = ShouldInstrument;
514 return ShouldInstrument;
517 std::string GCOVProfiler::mangleName(const DICompileUnit *CU,
518 GCovFileType OutputType) {
519 bool Notes = OutputType == GCovFileType::GCNO;
521 if (NamedMDNode *GCov = M->getNamedMetadata("llvm.gcov")) {
522 for (int i = 0, e = GCov->getNumOperands(); i != e; ++i) {
523 MDNode *N = GCov->getOperand(i);
524 bool ThreeElement = N->getNumOperands() == 3;
525 if (!ThreeElement && N->getNumOperands() != 2)
527 if (dyn_cast<MDNode>(N->getOperand(ThreeElement ? 2 : 1)) != CU)
531 // These nodes have no mangling to apply, it's stored mangled in the
533 MDString *NotesFile = dyn_cast<MDString>(N->getOperand(0));
534 MDString *DataFile = dyn_cast<MDString>(N->getOperand(1));
535 if (!NotesFile || !DataFile)
537 return Notes ? NotesFile->getString() : DataFile->getString();
540 MDString *GCovFile = dyn_cast<MDString>(N->getOperand(0));
544 SmallString<128> Filename = GCovFile->getString();
545 sys::path::replace_extension(Filename, Notes ? "gcno" : "gcda");
546 return Filename.str();
550 SmallString<128> Filename = CU->getFilename();
551 sys::path::replace_extension(Filename, Notes ? "gcno" : "gcda");
552 StringRef FName = sys::path::filename(Filename);
553 SmallString<128> CurPath;
554 if (sys::fs::current_path(CurPath)) return FName;
555 sys::path::append(CurPath, FName);
556 return CurPath.str();
559 bool GCOVProfiler::runOnModule(Module &M, const TargetLibraryInfo &TLI) {
562 Ctx = &M.getContext();
564 AddFlushBeforeForkAndExec();
566 FilterRe = createRegexesFromString(Options.Filter);
567 ExcludeRe = createRegexesFromString(Options.Exclude);
569 if (Options.EmitNotes) emitProfileNotes();
570 if (Options.EmitData) return emitProfileArcs();
574 PreservedAnalyses GCOVProfilerPass::run(Module &M,
575 ModuleAnalysisManager &AM) {
577 GCOVProfiler Profiler(GCOVOpts);
579 auto &TLI = AM.getResult<TargetLibraryAnalysis>(M);
580 if (!Profiler.runOnModule(M, TLI))
581 return PreservedAnalyses::all();
583 return PreservedAnalyses::none();
586 static bool functionHasLines(Function &F) {
587 // Check whether this function actually has any source lines. Not only
588 // do these waste space, they also can crash gcov.
591 // Debug intrinsic locations correspond to the location of the
592 // declaration, not necessarily any statements or expressions.
593 if (isa<DbgInfoIntrinsic>(&I)) continue;
595 const DebugLoc &Loc = I.getDebugLoc();
599 // Artificial lines such as calls to the global constructors.
600 if (Loc.getLine() == 0) continue;
608 static bool isUsingScopeBasedEH(Function &F) {
609 if (!F.hasPersonalityFn()) return false;
611 EHPersonality Personality = classifyEHPersonality(F.getPersonalityFn());
612 return isScopedEHPersonality(Personality);
615 static bool shouldKeepInEntry(BasicBlock::iterator It) {
616 if (isa<AllocaInst>(*It)) return true;
617 if (isa<DbgInfoIntrinsic>(*It)) return true;
618 if (auto *II = dyn_cast<IntrinsicInst>(It)) {
619 if (II->getIntrinsicID() == llvm::Intrinsic::localescape) return true;
625 void GCOVProfiler::AddFlushBeforeForkAndExec() {
626 SmallVector<Instruction *, 2> ForkAndExecs;
627 for (auto &F : M->functions()) {
628 for (auto &I : instructions(F)) {
629 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
630 if (Function *Callee = CI->getCalledFunction()) {
632 if (TLI->getLibFunc(*Callee, LF) &&
633 (LF == LibFunc_fork || LF == LibFunc_execl ||
634 LF == LibFunc_execle || LF == LibFunc_execlp ||
635 LF == LibFunc_execv || LF == LibFunc_execvp ||
636 LF == LibFunc_execve || LF == LibFunc_execvpe ||
637 LF == LibFunc_execvP)) {
638 ForkAndExecs.push_back(&I);
645 // We need to split the block after the fork/exec call
646 // because else the counters for the lines after will be
647 // the same as before the call.
648 for (auto I : ForkAndExecs) {
649 IRBuilder<> Builder(I);
650 FunctionType *FTy = FunctionType::get(Builder.getVoidTy(), {}, false);
651 Constant *GCOVFlush = M->getOrInsertFunction("__gcov_flush", FTy);
652 Builder.CreateCall(GCOVFlush);
653 I->getParent()->splitBasicBlock(I);
657 void GCOVProfiler::emitProfileNotes() {
658 NamedMDNode *CU_Nodes = M->getNamedMetadata("llvm.dbg.cu");
659 if (!CU_Nodes) return;
661 for (unsigned i = 0, e = CU_Nodes->getNumOperands(); i != e; ++i) {
662 // Each compile unit gets its own .gcno file. This means that whether we run
663 // this pass over the original .o's as they're produced, or run it after
664 // LTO, we'll generate the same .gcno files.
666 auto *CU = cast<DICompileUnit>(CU_Nodes->getOperand(i));
668 // Skip module skeleton (and module) CUs.
673 raw_fd_ostream out(mangleName(CU, GCovFileType::GCNO), EC, sys::fs::F_None);
675 Ctx->emitError(Twine("failed to open coverage notes file for writing: ") +
680 std::string EdgeDestinations;
682 unsigned FunctionIdent = 0;
683 for (auto &F : M->functions()) {
684 DISubprogram *SP = F.getSubprogram();
686 if (!functionHasLines(F) || !isFunctionInstrumented(F))
688 // TODO: Functions using scope-based EH are currently not supported.
689 if (isUsingScopeBasedEH(F)) continue;
691 // gcov expects every function to start with an entry block that has a
692 // single successor, so split the entry block to make sure of that.
693 BasicBlock &EntryBlock = F.getEntryBlock();
694 BasicBlock::iterator It = EntryBlock.begin();
695 while (shouldKeepInEntry(It))
697 EntryBlock.splitBasicBlock(It);
699 Funcs.push_back(make_unique<GCOVFunction>(SP, &F, &out, FunctionIdent++,
700 Options.UseCfgChecksum,
701 Options.ExitBlockBeforeBody));
702 GCOVFunction &Func = *Funcs.back();
704 // Add the function line number to the lines of the entry block
705 // to have a counter for the function definition.
706 uint32_t Line = SP->getLine();
707 auto Filename = getFilename(SP);
708 Func.getBlock(&EntryBlock).getFile(Filename).addLine(Line);
711 GCOVBlock &Block = Func.getBlock(&BB);
712 Instruction *TI = BB.getTerminator();
713 if (int successors = TI->getNumSuccessors()) {
714 for (int i = 0; i != successors; ++i) {
715 Block.addEdge(Func.getBlock(TI->getSuccessor(i)));
717 } else if (isa<ReturnInst>(TI)) {
718 Block.addEdge(Func.getReturnBlock());
722 // Debug intrinsic locations correspond to the location of the
723 // declaration, not necessarily any statements or expressions.
724 if (isa<DbgInfoIntrinsic>(&I)) continue;
726 const DebugLoc &Loc = I.getDebugLoc();
730 // Artificial lines such as calls to the global constructors.
731 if (Loc.getLine() == 0 || Loc.isImplicitCode())
734 if (Line == Loc.getLine()) continue;
735 Line = Loc.getLine();
736 if (SP != getDISubprogram(Loc.getScope()))
739 GCOVLines &Lines = Block.getFile(Filename);
740 Lines.addLine(Loc.getLine());
744 EdgeDestinations += Func.getEdgeDestinations();
747 FileChecksums.push_back(hash_value(EdgeDestinations));
748 out.write("oncg", 4);
749 out.write(ReversedVersion, 4);
750 out.write(reinterpret_cast<char*>(&FileChecksums.back()), 4);
752 for (auto &Func : Funcs) {
753 Func->setCfgChecksum(FileChecksums.back());
757 out.write("\0\0\0\0\0\0\0\0", 8); // EOF
762 bool GCOVProfiler::emitProfileArcs() {
763 NamedMDNode *CU_Nodes = M->getNamedMetadata("llvm.dbg.cu");
764 if (!CU_Nodes) return false;
767 for (unsigned i = 0, e = CU_Nodes->getNumOperands(); i != e; ++i) {
768 SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> CountersBySP;
769 for (auto &F : M->functions()) {
770 DISubprogram *SP = F.getSubprogram();
772 if (!functionHasLines(F) || !isFunctionInstrumented(F))
774 // TODO: Functions using scope-based EH are currently not supported.
775 if (isUsingScopeBasedEH(F)) continue;
776 if (!Result) Result = true;
778 DenseMap<std::pair<BasicBlock *, BasicBlock *>, unsigned> EdgeToCounter;
781 Instruction *TI = BB.getTerminator();
782 if (isa<ReturnInst>(TI)) {
783 EdgeToCounter[{&BB, nullptr}] = Edges++;
785 for (BasicBlock *Succ : successors(TI)) {
786 EdgeToCounter[{&BB, Succ}] = Edges++;
791 ArrayType *CounterTy =
792 ArrayType::get(Type::getInt64Ty(*Ctx), Edges);
793 GlobalVariable *Counters =
794 new GlobalVariable(*M, CounterTy, false,
795 GlobalValue::InternalLinkage,
796 Constant::getNullValue(CounterTy),
798 CountersBySP.push_back(std::make_pair(Counters, SP));
800 // If a BB has several predecessors, use a PHINode to select
801 // the correct counter.
803 const unsigned EdgeCount =
804 std::distance(pred_begin(&BB), pred_end(&BB));
806 // The phi node must be at the begin of the BB.
807 IRBuilder<> BuilderForPhi(&*BB.begin());
808 Type *Int64PtrTy = Type::getInt64PtrTy(*Ctx);
809 PHINode *Phi = BuilderForPhi.CreatePHI(Int64PtrTy, EdgeCount);
810 for (BasicBlock *Pred : predecessors(&BB)) {
811 auto It = EdgeToCounter.find({Pred, &BB});
812 assert(It != EdgeToCounter.end());
813 const unsigned Edge = It->second;
815 BuilderForPhi.CreateConstInBoundsGEP2_64(Counters, 0, Edge);
816 Phi->addIncoming(EdgeCounter, Pred);
819 // Skip phis, landingpads.
820 IRBuilder<> Builder(&*BB.getFirstInsertionPt());
821 Value *Count = Builder.CreateLoad(Phi);
822 Count = Builder.CreateAdd(Count, Builder.getInt64(1));
823 Builder.CreateStore(Count, Phi);
825 Instruction *TI = BB.getTerminator();
826 if (isa<ReturnInst>(TI)) {
827 auto It = EdgeToCounter.find({&BB, nullptr});
828 assert(It != EdgeToCounter.end());
829 const unsigned Edge = It->second;
831 Builder.CreateConstInBoundsGEP2_64(Counters, 0, Edge);
832 Value *Count = Builder.CreateLoad(Counter);
833 Count = Builder.CreateAdd(Count, Builder.getInt64(1));
834 Builder.CreateStore(Count, Counter);
840 Function *WriteoutF = insertCounterWriteout(CountersBySP);
841 Function *FlushF = insertFlush(CountersBySP);
843 // Create a small bit of code that registers the "__llvm_gcov_writeout" to
844 // be executed at exit and the "__llvm_gcov_flush" function to be executed
845 // when "__gcov_flush" is called.
846 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
847 Function *F = Function::Create(FTy, GlobalValue::InternalLinkage,
848 "__llvm_gcov_init", M);
849 F->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
850 F->setLinkage(GlobalValue::InternalLinkage);
851 F->addFnAttr(Attribute::NoInline);
852 if (Options.NoRedZone)
853 F->addFnAttr(Attribute::NoRedZone);
855 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
856 IRBuilder<> Builder(BB);
858 FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
860 PointerType::get(FTy, 0),
861 PointerType::get(FTy, 0)
863 FTy = FunctionType::get(Builder.getVoidTy(), Params, false);
865 // Initialize the environment and register the local writeout and flush
867 Constant *GCOVInit = M->getOrInsertFunction("llvm_gcov_init", FTy);
868 Builder.CreateCall(GCOVInit, {WriteoutF, FlushF});
869 Builder.CreateRetVoid();
871 appendToGlobalCtors(*M, F, 0);
877 Constant *GCOVProfiler::getStartFileFunc() {
879 Type::getInt8PtrTy(*Ctx), // const char *orig_filename
880 Type::getInt8PtrTy(*Ctx), // const char version[4]
881 Type::getInt32Ty(*Ctx), // uint32_t checksum
883 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), Args, false);
884 auto *Res = M->getOrInsertFunction("llvm_gcda_start_file", FTy);
885 if (Function *FunRes = dyn_cast<Function>(Res))
886 if (auto AK = TLI->getExtAttrForI32Param(false))
887 FunRes->addParamAttr(2, AK);
892 Constant *GCOVProfiler::getEmitFunctionFunc() {
894 Type::getInt32Ty(*Ctx), // uint32_t ident
895 Type::getInt8PtrTy(*Ctx), // const char *function_name
896 Type::getInt32Ty(*Ctx), // uint32_t func_checksum
897 Type::getInt8Ty(*Ctx), // uint8_t use_extra_checksum
898 Type::getInt32Ty(*Ctx), // uint32_t cfg_checksum
900 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), Args, false);
901 auto *Res = M->getOrInsertFunction("llvm_gcda_emit_function", FTy);
902 if (Function *FunRes = dyn_cast<Function>(Res))
903 if (auto AK = TLI->getExtAttrForI32Param(false)) {
904 FunRes->addParamAttr(0, AK);
905 FunRes->addParamAttr(2, AK);
906 FunRes->addParamAttr(3, AK);
907 FunRes->addParamAttr(4, AK);
912 Constant *GCOVProfiler::getEmitArcsFunc() {
914 Type::getInt32Ty(*Ctx), // uint32_t num_counters
915 Type::getInt64PtrTy(*Ctx), // uint64_t *counters
917 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), Args, false);
918 auto *Res = M->getOrInsertFunction("llvm_gcda_emit_arcs", FTy);
919 if (Function *FunRes = dyn_cast<Function>(Res))
920 if (auto AK = TLI->getExtAttrForI32Param(false))
921 FunRes->addParamAttr(0, AK);
925 Constant *GCOVProfiler::getSummaryInfoFunc() {
926 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
927 return M->getOrInsertFunction("llvm_gcda_summary_info", FTy);
930 Constant *GCOVProfiler::getEndFileFunc() {
931 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
932 return M->getOrInsertFunction("llvm_gcda_end_file", FTy);
935 Function *GCOVProfiler::insertCounterWriteout(
936 ArrayRef<std::pair<GlobalVariable *, MDNode *> > CountersBySP) {
937 FunctionType *WriteoutFTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
938 Function *WriteoutF = M->getFunction("__llvm_gcov_writeout");
940 WriteoutF = Function::Create(WriteoutFTy, GlobalValue::InternalLinkage,
941 "__llvm_gcov_writeout", M);
942 WriteoutF->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
943 WriteoutF->addFnAttr(Attribute::NoInline);
944 if (Options.NoRedZone)
945 WriteoutF->addFnAttr(Attribute::NoRedZone);
947 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", WriteoutF);
948 IRBuilder<> Builder(BB);
950 Constant *StartFile = getStartFileFunc();
951 Constant *EmitFunction = getEmitFunctionFunc();
952 Constant *EmitArcs = getEmitArcsFunc();
953 Constant *SummaryInfo = getSummaryInfoFunc();
954 Constant *EndFile = getEndFileFunc();
956 NamedMDNode *CUNodes = M->getNamedMetadata("llvm.dbg.cu");
958 Builder.CreateRetVoid();
962 // Collect the relevant data into a large constant data structure that we can
963 // walk to write out everything.
964 StructType *StartFileCallArgsTy = StructType::create(
965 {Builder.getInt8PtrTy(), Builder.getInt8PtrTy(), Builder.getInt32Ty()});
966 StructType *EmitFunctionCallArgsTy = StructType::create(
967 {Builder.getInt32Ty(), Builder.getInt8PtrTy(), Builder.getInt32Ty(),
968 Builder.getInt8Ty(), Builder.getInt32Ty()});
969 StructType *EmitArcsCallArgsTy = StructType::create(
970 {Builder.getInt32Ty(), Builder.getInt64Ty()->getPointerTo()});
971 StructType *FileInfoTy =
972 StructType::create({StartFileCallArgsTy, Builder.getInt32Ty(),
973 EmitFunctionCallArgsTy->getPointerTo(),
974 EmitArcsCallArgsTy->getPointerTo()});
976 Constant *Zero32 = Builder.getInt32(0);
977 // Build an explicit array of two zeros for use in ConstantExpr GEP building.
978 Constant *TwoZero32s[] = {Zero32, Zero32};
980 SmallVector<Constant *, 8> FileInfos;
981 for (int i : llvm::seq<int>(0, CUNodes->getNumOperands())) {
982 auto *CU = cast<DICompileUnit>(CUNodes->getOperand(i));
984 // Skip module skeleton (and module) CUs.
988 std::string FilenameGcda = mangleName(CU, GCovFileType::GCDA);
989 uint32_t CfgChecksum = FileChecksums.empty() ? 0 : FileChecksums[i];
990 auto *StartFileCallArgs = ConstantStruct::get(
991 StartFileCallArgsTy, {Builder.CreateGlobalStringPtr(FilenameGcda),
992 Builder.CreateGlobalStringPtr(ReversedVersion),
993 Builder.getInt32(CfgChecksum)});
995 SmallVector<Constant *, 8> EmitFunctionCallArgsArray;
996 SmallVector<Constant *, 8> EmitArcsCallArgsArray;
997 for (int j : llvm::seq<int>(0, CountersBySP.size())) {
998 auto *SP = cast_or_null<DISubprogram>(CountersBySP[j].second);
999 uint32_t FuncChecksum = Funcs.empty() ? 0 : Funcs[j]->getFuncChecksum();
1000 EmitFunctionCallArgsArray.push_back(ConstantStruct::get(
1001 EmitFunctionCallArgsTy,
1002 {Builder.getInt32(j),
1003 Options.FunctionNamesInData
1004 ? Builder.CreateGlobalStringPtr(getFunctionName(SP))
1005 : Constant::getNullValue(Builder.getInt8PtrTy()),
1006 Builder.getInt32(FuncChecksum),
1007 Builder.getInt8(Options.UseCfgChecksum),
1008 Builder.getInt32(CfgChecksum)}));
1010 GlobalVariable *GV = CountersBySP[j].first;
1011 unsigned Arcs = cast<ArrayType>(GV->getValueType())->getNumElements();
1012 EmitArcsCallArgsArray.push_back(ConstantStruct::get(
1014 {Builder.getInt32(Arcs), ConstantExpr::getInBoundsGetElementPtr(
1015 GV->getValueType(), GV, TwoZero32s)}));
1017 // Create global arrays for the two emit calls.
1018 int CountersSize = CountersBySP.size();
1019 assert(CountersSize == (int)EmitFunctionCallArgsArray.size() &&
1020 "Mismatched array size!");
1021 assert(CountersSize == (int)EmitArcsCallArgsArray.size() &&
1022 "Mismatched array size!");
1023 auto *EmitFunctionCallArgsArrayTy =
1024 ArrayType::get(EmitFunctionCallArgsTy, CountersSize);
1025 auto *EmitFunctionCallArgsArrayGV = new GlobalVariable(
1026 *M, EmitFunctionCallArgsArrayTy, /*isConstant*/ true,
1027 GlobalValue::InternalLinkage,
1028 ConstantArray::get(EmitFunctionCallArgsArrayTy,
1029 EmitFunctionCallArgsArray),
1030 Twine("__llvm_internal_gcov_emit_function_args.") + Twine(i));
1031 auto *EmitArcsCallArgsArrayTy =
1032 ArrayType::get(EmitArcsCallArgsTy, CountersSize);
1033 EmitFunctionCallArgsArrayGV->setUnnamedAddr(
1034 GlobalValue::UnnamedAddr::Global);
1035 auto *EmitArcsCallArgsArrayGV = new GlobalVariable(
1036 *M, EmitArcsCallArgsArrayTy, /*isConstant*/ true,
1037 GlobalValue::InternalLinkage,
1038 ConstantArray::get(EmitArcsCallArgsArrayTy, EmitArcsCallArgsArray),
1039 Twine("__llvm_internal_gcov_emit_arcs_args.") + Twine(i));
1040 EmitArcsCallArgsArrayGV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1042 FileInfos.push_back(ConstantStruct::get(
1044 {StartFileCallArgs, Builder.getInt32(CountersSize),
1045 ConstantExpr::getInBoundsGetElementPtr(EmitFunctionCallArgsArrayTy,
1046 EmitFunctionCallArgsArrayGV,
1048 ConstantExpr::getInBoundsGetElementPtr(
1049 EmitArcsCallArgsArrayTy, EmitArcsCallArgsArrayGV, TwoZero32s)}));
1052 // If we didn't find anything to actually emit, bail on out.
1053 if (FileInfos.empty()) {
1054 Builder.CreateRetVoid();
1058 // To simplify code, we cap the number of file infos we write out to fit
1059 // easily in a 32-bit signed integer. This gives consistent behavior between
1060 // 32-bit and 64-bit systems without requiring (potentially very slow) 64-bit
1061 // operations on 32-bit systems. It also seems unreasonable to try to handle
1062 // more than 2 billion files.
1063 if ((int64_t)FileInfos.size() > (int64_t)INT_MAX)
1064 FileInfos.resize(INT_MAX);
1066 // Create a global for the entire data structure so we can walk it more
1068 auto *FileInfoArrayTy = ArrayType::get(FileInfoTy, FileInfos.size());
1069 auto *FileInfoArrayGV = new GlobalVariable(
1070 *M, FileInfoArrayTy, /*isConstant*/ true, GlobalValue::InternalLinkage,
1071 ConstantArray::get(FileInfoArrayTy, FileInfos),
1072 "__llvm_internal_gcov_emit_file_info");
1073 FileInfoArrayGV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1075 // Create the CFG for walking this data structure.
1076 auto *FileLoopHeader =
1077 BasicBlock::Create(*Ctx, "file.loop.header", WriteoutF);
1078 auto *CounterLoopHeader =
1079 BasicBlock::Create(*Ctx, "counter.loop.header", WriteoutF);
1080 auto *FileLoopLatch = BasicBlock::Create(*Ctx, "file.loop.latch", WriteoutF);
1081 auto *ExitBB = BasicBlock::Create(*Ctx, "exit", WriteoutF);
1083 // We always have at least one file, so just branch to the header.
1084 Builder.CreateBr(FileLoopHeader);
1086 // The index into the files structure is our loop induction variable.
1087 Builder.SetInsertPoint(FileLoopHeader);
1089 Builder.CreatePHI(Builder.getInt32Ty(), /*NumReservedValues*/ 2);
1090 IV->addIncoming(Builder.getInt32(0), BB);
1092 Builder.CreateInBoundsGEP(FileInfoArrayGV, {Builder.getInt32(0), IV});
1093 auto *StartFileCallArgsPtr = Builder.CreateStructGEP(FileInfoPtr, 0);
1094 auto *StartFileCall = Builder.CreateCall(
1096 {Builder.CreateLoad(Builder.CreateStructGEP(StartFileCallArgsPtr, 0)),
1097 Builder.CreateLoad(Builder.CreateStructGEP(StartFileCallArgsPtr, 1)),
1098 Builder.CreateLoad(Builder.CreateStructGEP(StartFileCallArgsPtr, 2))});
1099 if (auto AK = TLI->getExtAttrForI32Param(false))
1100 StartFileCall->addParamAttr(2, AK);
1102 Builder.CreateLoad(Builder.CreateStructGEP(FileInfoPtr, 1));
1103 auto *EmitFunctionCallArgsArray =
1104 Builder.CreateLoad(Builder.CreateStructGEP(FileInfoPtr, 2));
1105 auto *EmitArcsCallArgsArray =
1106 Builder.CreateLoad(Builder.CreateStructGEP(FileInfoPtr, 3));
1107 auto *EnterCounterLoopCond =
1108 Builder.CreateICmpSLT(Builder.getInt32(0), NumCounters);
1109 Builder.CreateCondBr(EnterCounterLoopCond, CounterLoopHeader, FileLoopLatch);
1111 Builder.SetInsertPoint(CounterLoopHeader);
1112 auto *JV = Builder.CreatePHI(Builder.getInt32Ty(), /*NumReservedValues*/ 2);
1113 JV->addIncoming(Builder.getInt32(0), FileLoopHeader);
1114 auto *EmitFunctionCallArgsPtr =
1115 Builder.CreateInBoundsGEP(EmitFunctionCallArgsArray, {JV});
1116 auto *EmitFunctionCall = Builder.CreateCall(
1118 {Builder.CreateLoad(Builder.CreateStructGEP(EmitFunctionCallArgsPtr, 0)),
1119 Builder.CreateLoad(Builder.CreateStructGEP(EmitFunctionCallArgsPtr, 1)),
1120 Builder.CreateLoad(Builder.CreateStructGEP(EmitFunctionCallArgsPtr, 2)),
1121 Builder.CreateLoad(Builder.CreateStructGEP(EmitFunctionCallArgsPtr, 3)),
1123 Builder.CreateStructGEP(EmitFunctionCallArgsPtr, 4))});
1124 if (auto AK = TLI->getExtAttrForI32Param(false)) {
1125 EmitFunctionCall->addParamAttr(0, AK);
1126 EmitFunctionCall->addParamAttr(2, AK);
1127 EmitFunctionCall->addParamAttr(3, AK);
1128 EmitFunctionCall->addParamAttr(4, AK);
1130 auto *EmitArcsCallArgsPtr =
1131 Builder.CreateInBoundsGEP(EmitArcsCallArgsArray, {JV});
1132 auto *EmitArcsCall = Builder.CreateCall(
1134 {Builder.CreateLoad(Builder.CreateStructGEP(EmitArcsCallArgsPtr, 0)),
1135 Builder.CreateLoad(Builder.CreateStructGEP(EmitArcsCallArgsPtr, 1))});
1136 if (auto AK = TLI->getExtAttrForI32Param(false))
1137 EmitArcsCall->addParamAttr(0, AK);
1138 auto *NextJV = Builder.CreateAdd(JV, Builder.getInt32(1));
1139 auto *CounterLoopCond = Builder.CreateICmpSLT(NextJV, NumCounters);
1140 Builder.CreateCondBr(CounterLoopCond, CounterLoopHeader, FileLoopLatch);
1141 JV->addIncoming(NextJV, CounterLoopHeader);
1143 Builder.SetInsertPoint(FileLoopLatch);
1144 Builder.CreateCall(SummaryInfo, {});
1145 Builder.CreateCall(EndFile, {});
1146 auto *NextIV = Builder.CreateAdd(IV, Builder.getInt32(1));
1147 auto *FileLoopCond =
1148 Builder.CreateICmpSLT(NextIV, Builder.getInt32(FileInfos.size()));
1149 Builder.CreateCondBr(FileLoopCond, FileLoopHeader, ExitBB);
1150 IV->addIncoming(NextIV, FileLoopLatch);
1152 Builder.SetInsertPoint(ExitBB);
1153 Builder.CreateRetVoid();
1158 Function *GCOVProfiler::
1159 insertFlush(ArrayRef<std::pair<GlobalVariable*, MDNode*> > CountersBySP) {
1160 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
1161 Function *FlushF = M->getFunction("__llvm_gcov_flush");
1163 FlushF = Function::Create(FTy, GlobalValue::InternalLinkage,
1164 "__llvm_gcov_flush", M);
1166 FlushF->setLinkage(GlobalValue::InternalLinkage);
1167 FlushF->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1168 FlushF->addFnAttr(Attribute::NoInline);
1169 if (Options.NoRedZone)
1170 FlushF->addFnAttr(Attribute::NoRedZone);
1172 BasicBlock *Entry = BasicBlock::Create(*Ctx, "entry", FlushF);
1174 // Write out the current counters.
1175 Constant *WriteoutF = M->getFunction("__llvm_gcov_writeout");
1176 assert(WriteoutF && "Need to create the writeout function first!");
1178 IRBuilder<> Builder(Entry);
1179 Builder.CreateCall(WriteoutF, {});
1181 // Zero out the counters.
1182 for (const auto &I : CountersBySP) {
1183 GlobalVariable *GV = I.first;
1184 Constant *Null = Constant::getNullValue(GV->getValueType());
1185 Builder.CreateStore(Null, GV);
1188 Type *RetTy = FlushF->getReturnType();
1189 if (RetTy == Type::getVoidTy(*Ctx))
1190 Builder.CreateRetVoid();
1191 else if (RetTy->isIntegerTy())
1192 // Used if __llvm_gcov_flush was implicitly declared.
1193 Builder.CreateRet(ConstantInt::get(RetTy, 0));
1195 report_fatal_error("invalid return type for __llvm_gcov_flush");