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 #define DEBUG_TYPE "insert-gcov-profiling"
19 #include "ProfilingUtils.h"
20 #include "llvm/Transforms/Instrumentation.h"
21 #include "llvm/Analysis/DebugInfo.h"
22 #include "llvm/Module.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/DebugLoc.h"
28 #include "llvm/Support/InstIterator.h"
29 #include "llvm/Support/IRBuilder.h"
30 #include "llvm/Support/PathV2.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/ADT/StringMap.h"
36 #include "llvm/ADT/UniqueVector.h"
42 class GCOVProfiler : public ModulePass {
46 : ModulePass(ID), EmitNotes(true), EmitData(true), Use402Format(false) {
47 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
49 GCOVProfiler(bool EmitNotes, bool EmitData, bool use402Format = false)
50 : ModulePass(ID), EmitNotes(EmitNotes), EmitData(EmitData),
51 Use402Format(use402Format) {
52 assert((EmitNotes || EmitData) && "GCOVProfiler asked to do nothing?");
53 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
55 virtual const char *getPassName() const {
56 return "GCOV Profiler";
60 bool runOnModule(Module &M);
62 // Create the GCNO files for the Module based on DebugInfo.
63 void emitGCNO(DebugInfoFinder &DIF);
65 // Modify the program to track transitions along edges and call into the
66 // profiling runtime to emit .gcda files when run.
67 bool emitProfileArcs(DebugInfoFinder &DIF);
69 // Get pointers to the functions in the runtime library.
70 Constant *getStartFileFunc();
71 Constant *getIncrementIndirectCounterFunc();
72 Constant *getEmitFunctionFunc();
73 Constant *getEmitArcsFunc();
74 Constant *getEndFileFunc();
76 // Create or retrieve an i32 state value that is used to represent the
77 // pred block number for certain non-trivial edges.
78 GlobalVariable *getEdgeStateValue();
80 // Produce a table of pointers to counters, by predecessor and successor
82 GlobalVariable *buildEdgeLookupTable(Function *F,
83 GlobalVariable *Counter,
84 const UniqueVector<BasicBlock *> &Preds,
85 const UniqueVector<BasicBlock *> &Succs);
87 // Add the function to write out all our counters to the global destructor
89 void insertCounterWriteout(DebugInfoFinder &,
90 SmallVector<std::pair<GlobalVariable *,
93 std::string mangleName(DICompileUnit CU, std::string NewStem);
104 char GCOVProfiler::ID = 0;
105 INITIALIZE_PASS(GCOVProfiler, "insert-gcov-profiling",
106 "Insert instrumentation for GCOV profiling", false, false)
108 ModulePass *llvm::createGCOVProfilerPass(bool EmitNotes, bool EmitData,
110 return new GCOVProfiler(EmitNotes, EmitData, Use402Format);
113 static DISubprogram findSubprogram(DIScope Scope) {
114 while (!Scope.isSubprogram()) {
115 assert(Scope.isLexicalBlock() &&
116 "Debug location not lexical block or subprogram");
117 Scope = DILexicalBlock(Scope).getContext();
119 return DISubprogram(Scope);
125 static const char *LinesTag;
126 static const char *FunctionTag;
127 static const char *BlockTag;
128 static const char *EdgeTag;
132 void writeBytes(const char *Bytes, int Size) {
133 os->write(Bytes, Size);
136 void write(uint32_t i) {
137 writeBytes(reinterpret_cast<char*>(&i), 4);
140 // Returns the length measured in 4-byte blocks that will be used to
141 // represent this string in a GCOV file
142 unsigned lengthOfGCOVString(StringRef s) {
143 // A GCOV string is a length, followed by a NUL, then between 0 and 3 NULs
144 // padding out to the next 4-byte word. The length is measured in 4-byte
145 // words including padding, not bytes of actual string.
146 return (s.size() / 4) + 1;
149 void writeGCOVString(StringRef s) {
150 uint32_t Len = lengthOfGCOVString(s);
152 writeBytes(s.data(), s.size());
154 // Write 1 to 4 bytes of NUL padding.
155 assert((unsigned)(4 - (s.size() % 4)) > 0);
156 assert((unsigned)(4 - (s.size() % 4)) <= 4);
157 writeBytes("\0\0\0\0", 4 - (s.size() % 4));
162 const char *GCOVRecord::LinesTag = "\0\0\x45\x01";
163 const char *GCOVRecord::FunctionTag = "\0\0\0\1";
164 const char *GCOVRecord::BlockTag = "\0\0\x41\x01";
165 const char *GCOVRecord::EdgeTag = "\0\0\x43\x01";
170 // Constructed only by requesting it from a GCOVBlock, this object stores a
171 // list of line numbers and a single filename, representing lines that belong
173 class GCOVLines : public GCOVRecord {
175 void addLine(uint32_t Line) {
176 Lines.push_back(Line);
180 return lengthOfGCOVString(Filename) + 2 + Lines.size();
184 friend class GCOVBlock;
186 GCOVLines(std::string Filename, raw_ostream *os)
187 : Filename(Filename) {
191 std::string Filename;
192 SmallVector<uint32_t, 32> Lines;
195 // Represent a basic block in GCOV. Each block has a unique number in the
196 // function, number of lines belonging to each block, and a set of edges to
198 class GCOVBlock : public GCOVRecord {
200 GCOVLines &getFile(std::string Filename) {
201 GCOVLines *&Lines = LinesByFile[Filename];
203 Lines = new GCOVLines(Filename, os);
208 void addEdge(GCOVBlock &Successor) {
209 OutEdges.push_back(&Successor);
214 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
215 E = LinesByFile.end(); I != E; ++I) {
216 Len += I->second->length();
219 writeBytes(LinesTag, 4);
222 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
223 E = LinesByFile.end(); I != E; ++I) {
225 writeGCOVString(I->second->Filename);
226 for (int i = 0, e = I->second->Lines.size(); i != e; ++i) {
227 write(I->second->Lines[i]);
235 DeleteContainerSeconds(LinesByFile);
239 friend class GCOVFunction;
241 GCOVBlock(uint32_t Number, raw_ostream *os)
247 StringMap<GCOVLines *> LinesByFile;
248 SmallVector<GCOVBlock *, 4> OutEdges;
251 // A function has a unique identifier, a checksum (we leave as zero) and a
252 // set of blocks and a map of edges between blocks. This is the only GCOV
253 // object users can construct, the blocks and lines will be rooted here.
254 class GCOVFunction : public GCOVRecord {
256 GCOVFunction(DISubprogram SP, raw_ostream *os, bool Use402Format) {
259 Function *F = SP.getFunction();
261 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
262 Blocks[BB] = new GCOVBlock(i++, os);
264 ReturnBlock = new GCOVBlock(i++, os);
266 writeBytes(FunctionTag, 4);
267 uint32_t BlockLen = 1 + 1 + 1 + lengthOfGCOVString(SP.getName()) +
268 1 + lengthOfGCOVString(SP.getFilename()) + 1;
270 ++BlockLen; // For second checksum.
272 uint32_t Ident = reinterpret_cast<intptr_t>((MDNode*)SP);
274 write(0); // checksum #1
276 write(0); // checksum #2
277 writeGCOVString(SP.getName());
278 writeGCOVString(SP.getFilename());
279 write(SP.getLineNumber());
283 DeleteContainerSeconds(Blocks);
287 GCOVBlock &getBlock(BasicBlock *BB) {
291 GCOVBlock &getReturnBlock() {
296 // Emit count of blocks.
297 writeBytes(BlockTag, 4);
298 write(Blocks.size() + 1);
299 for (int i = 0, e = Blocks.size() + 1; i != e; ++i) {
300 write(0); // No flags on our blocks.
303 // Emit edges between blocks.
304 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
305 E = Blocks.end(); I != E; ++I) {
306 GCOVBlock &Block = *I->second;
307 if (Block.OutEdges.empty()) continue;
309 writeBytes(EdgeTag, 4);
310 write(Block.OutEdges.size() * 2 + 1);
312 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i) {
313 write(Block.OutEdges[i]->Number);
314 write(0); // no flags
318 // Emit lines for each block.
319 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
320 E = Blocks.end(); I != E; ++I) {
321 I->second->writeOut();
326 DenseMap<BasicBlock *, GCOVBlock *> Blocks;
327 GCOVBlock *ReturnBlock;
331 std::string GCOVProfiler::mangleName(DICompileUnit CU, std::string NewStem) {
332 if (NamedMDNode *GCov = M->getNamedMetadata("llvm.gcov")) {
333 for (int i = 0, e = GCov->getNumOperands(); i != e; ++i) {
334 MDNode *N = GCov->getOperand(i);
335 if (N->getNumOperands() != 2) continue;
336 MDString *GCovFile = dyn_cast<MDString>(N->getOperand(0));
337 MDNode *CompileUnit = dyn_cast<MDNode>(N->getOperand(1));
338 if (!GCovFile || !CompileUnit) continue;
339 if (CompileUnit == CU) {
340 SmallString<128> Filename = GCovFile->getString();
341 sys::path::replace_extension(Filename, NewStem);
342 return Filename.str();
347 SmallString<128> Filename = CU.getFilename();
348 sys::path::replace_extension(Filename, NewStem);
349 return sys::path::filename(Filename.str());
352 bool GCOVProfiler::runOnModule(Module &M) {
354 Ctx = &M.getContext();
357 DIF.processModule(M);
359 if (EmitNotes) emitGCNO(DIF);
360 if (EmitData) return emitProfileArcs(DIF);
364 void GCOVProfiler::emitGCNO(DebugInfoFinder &DIF) {
365 DenseMap<const MDNode *, raw_fd_ostream *> GcnoFiles;
366 for (DebugInfoFinder::iterator I = DIF.compile_unit_begin(),
367 E = DIF.compile_unit_end(); I != E; ++I) {
368 // Each compile unit gets its own .gcno file. This means that whether we run
369 // this pass over the original .o's as they're produced, or run it after
370 // LTO, we'll generate the same .gcno files.
372 DICompileUnit CU(*I);
373 raw_fd_ostream *&out = GcnoFiles[CU];
374 std::string ErrorInfo;
375 out = new raw_fd_ostream(mangleName(CU, "gcno").c_str(), ErrorInfo,
376 raw_fd_ostream::F_Binary);
378 out->write("oncg*404MVLL", 12);
380 out->write("oncg*204MVLL", 12);
383 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
384 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
385 DISubprogram SP(*SPI);
386 raw_fd_ostream *&os = GcnoFiles[SP.getCompileUnit()];
388 Function *F = SP.getFunction();
390 GCOVFunction Func(SP, os, Use402Format);
392 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
393 GCOVBlock &Block = Func.getBlock(BB);
394 TerminatorInst *TI = BB->getTerminator();
395 if (int successors = TI->getNumSuccessors()) {
396 for (int i = 0; i != successors; ++i) {
397 Block.addEdge(Func.getBlock(TI->getSuccessor(i)));
399 } else if (isa<ReturnInst>(TI)) {
400 Block.addEdge(Func.getReturnBlock());
404 for (BasicBlock::iterator I = BB->begin(), IE = BB->end(); I != IE; ++I) {
405 const DebugLoc &Loc = I->getDebugLoc();
406 if (Loc.isUnknown()) continue;
407 if (Line == Loc.getLine()) continue;
408 Line = Loc.getLine();
409 if (SP != findSubprogram(DIScope(Loc.getScope(*Ctx)))) continue;
411 GCOVLines &Lines = Block.getFile(SP.getFilename());
412 Lines.addLine(Loc.getLine());
418 for (DenseMap<const MDNode *, raw_fd_ostream *>::iterator
419 I = GcnoFiles.begin(), E = GcnoFiles.end(); I != E; ++I) {
420 raw_fd_ostream *&out = I->second;
421 out->write("\0\0\0\0\0\0\0\0", 8); // EOF
427 bool GCOVProfiler::emitProfileArcs(DebugInfoFinder &DIF) {
428 if (DIF.subprogram_begin() == DIF.subprogram_end())
431 SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> CountersBySP;
432 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
433 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
434 DISubprogram SP(*SPI);
435 Function *F = SP.getFunction();
439 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
440 TerminatorInst *TI = BB->getTerminator();
441 if (isa<ReturnInst>(TI))
444 Edges += TI->getNumSuccessors();
447 ArrayType *CounterTy =
448 ArrayType::get(Type::getInt64Ty(*Ctx), Edges);
449 GlobalVariable *Counters =
450 new GlobalVariable(*M, CounterTy, false,
451 GlobalValue::InternalLinkage,
452 Constant::getNullValue(CounterTy),
453 "__llvm_gcov_ctr", 0, false, 0);
454 CountersBySP.push_back(std::make_pair(Counters, (MDNode*)SP));
456 UniqueVector<BasicBlock *> ComplexEdgePreds;
457 UniqueVector<BasicBlock *> ComplexEdgeSuccs;
460 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
461 TerminatorInst *TI = BB->getTerminator();
462 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
464 IRBuilder<> Builder(TI);
466 if (Successors == 1) {
467 Value *Counter = Builder.CreateConstInBoundsGEP2_64(Counters, 0,
469 Value *Count = Builder.CreateLoad(Counter);
470 Count = Builder.CreateAdd(Count,
471 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
472 Builder.CreateStore(Count, Counter);
473 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
474 Value *Sel = Builder.CreateSelect(
476 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge),
477 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge + 1));
478 SmallVector<Value *, 2> Idx;
479 Idx.push_back(Constant::getNullValue(Type::getInt64Ty(*Ctx)));
481 Value *Counter = Builder.CreateInBoundsGEP(Counters, Idx);
482 Value *Count = Builder.CreateLoad(Counter);
483 Count = Builder.CreateAdd(Count,
484 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
485 Builder.CreateStore(Count, Counter);
487 ComplexEdgePreds.insert(BB);
488 for (int i = 0; i != Successors; ++i)
489 ComplexEdgeSuccs.insert(TI->getSuccessor(i));
495 if (!ComplexEdgePreds.empty()) {
496 GlobalVariable *EdgeTable =
497 buildEdgeLookupTable(F, Counters,
498 ComplexEdgePreds, ComplexEdgeSuccs);
499 GlobalVariable *EdgeState = getEdgeStateValue();
501 Type *Int32Ty = Type::getInt32Ty(*Ctx);
502 for (int i = 0, e = ComplexEdgePreds.size(); i != e; ++i) {
503 IRBuilder<> Builder(ComplexEdgePreds[i+1]->getTerminator());
504 Builder.CreateStore(ConstantInt::get(Int32Ty, i), EdgeState);
506 for (int i = 0, e = ComplexEdgeSuccs.size(); i != e; ++i) {
507 // call runtime to perform increment
508 BasicBlock::iterator InsertPt = ComplexEdgeSuccs[i+1]->getFirstNonPHI();
509 if (isa<LandingPadInst>(InsertPt)) ++InsertPt;
510 IRBuilder<> Builder(InsertPt);
511 Value *CounterPtrArray =
512 Builder.CreateConstInBoundsGEP2_64(EdgeTable, 0,
513 i * ComplexEdgePreds.size());
514 Builder.CreateCall2(getIncrementIndirectCounterFunc(),
515 EdgeState, CounterPtrArray);
516 // clear the predecessor number
517 Builder.CreateStore(ConstantInt::get(Int32Ty, 0xffffffff), EdgeState);
522 insertCounterWriteout(DIF, CountersBySP);
527 // All edges with successors that aren't branches are "complex", because it
528 // requires complex logic to pick which counter to update.
529 GlobalVariable *GCOVProfiler::buildEdgeLookupTable(
531 GlobalVariable *Counters,
532 const UniqueVector<BasicBlock *> &Preds,
533 const UniqueVector<BasicBlock *> &Succs) {
534 // TODO: support invoke, threads. We rely on the fact that nothing can modify
535 // the whole-Module pred edge# between the time we set it and the time we next
536 // read it. Threads and invoke make this untrue.
538 // emit [(succs * preds) x i64*], logically [succ x [pred x i64*]].
539 Type *Int64PtrTy = Type::getInt64PtrTy(*Ctx);
540 ArrayType *EdgeTableTy = ArrayType::get(
541 Int64PtrTy, Succs.size() * Preds.size());
543 Constant **EdgeTable = new Constant*[Succs.size() * Preds.size()];
544 Constant *NullValue = Constant::getNullValue(Int64PtrTy);
545 for (int i = 0, ie = Succs.size() * Preds.size(); i != ie; ++i)
546 EdgeTable[i] = NullValue;
549 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
550 TerminatorInst *TI = BB->getTerminator();
551 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
552 if (Successors > 1 && !isa<BranchInst>(TI) && !isa<ReturnInst>(TI)) {
553 for (int i = 0; i != Successors; ++i) {
554 BasicBlock *Succ = TI->getSuccessor(i);
555 IRBuilder<> builder(Succ);
556 Value *Counter = builder.CreateConstInBoundsGEP2_64(Counters, 0,
558 EdgeTable[((Succs.idFor(Succ)-1) * Preds.size()) +
559 (Preds.idFor(BB)-1)] = cast<Constant>(Counter);
565 ArrayRef<Constant*> V(&EdgeTable[0], Succs.size() * Preds.size());
566 GlobalVariable *EdgeTableGV =
568 *M, EdgeTableTy, true, GlobalValue::InternalLinkage,
569 ConstantArray::get(EdgeTableTy, V),
570 "__llvm_gcda_edge_table");
571 EdgeTableGV->setUnnamedAddr(true);
575 Constant *GCOVProfiler::getStartFileFunc() {
576 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
577 Type::getInt8PtrTy(*Ctx), false);
578 return M->getOrInsertFunction("llvm_gcda_start_file", FTy);
581 Constant *GCOVProfiler::getIncrementIndirectCounterFunc() {
583 Type::getInt32PtrTy(*Ctx), // uint32_t *predecessor
584 Type::getInt64PtrTy(*Ctx)->getPointerTo(), // uint64_t **state_table_row
586 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
588 return M->getOrInsertFunction("llvm_gcda_increment_indirect_counter", FTy);
591 Constant *GCOVProfiler::getEmitFunctionFunc() {
593 Type::getInt32Ty(*Ctx), // uint32_t ident
594 Type::getInt8PtrTy(*Ctx), // const char *function_name
596 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
598 return M->getOrInsertFunction("llvm_gcda_emit_function", FTy);
601 Constant *GCOVProfiler::getEmitArcsFunc() {
603 Type::getInt32Ty(*Ctx), // uint32_t num_counters
604 Type::getInt64PtrTy(*Ctx), // uint64_t *counters
606 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
608 return M->getOrInsertFunction("llvm_gcda_emit_arcs", FTy);
611 Constant *GCOVProfiler::getEndFileFunc() {
612 FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
613 return M->getOrInsertFunction("llvm_gcda_end_file", FTy);
616 GlobalVariable *GCOVProfiler::getEdgeStateValue() {
617 GlobalVariable *GV = M->getGlobalVariable("__llvm_gcov_global_state_pred");
619 GV = new GlobalVariable(*M, Type::getInt32Ty(*Ctx), false,
620 GlobalValue::InternalLinkage,
621 ConstantInt::get(Type::getInt32Ty(*Ctx),
623 "__llvm_gcov_global_state_pred");
624 GV->setUnnamedAddr(true);
629 void GCOVProfiler::insertCounterWriteout(
630 DebugInfoFinder &DIF,
631 SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> &CountersBySP) {
632 FunctionType *WriteoutFTy =
633 FunctionType::get(Type::getVoidTy(*Ctx), false);
634 Function *WriteoutF = Function::Create(WriteoutFTy,
635 GlobalValue::InternalLinkage,
636 "__llvm_gcov_writeout", M);
637 WriteoutF->setUnnamedAddr(true);
638 BasicBlock *BB = BasicBlock::Create(*Ctx, "", WriteoutF);
639 IRBuilder<> Builder(BB);
641 Constant *StartFile = getStartFileFunc();
642 Constant *EmitFunction = getEmitFunctionFunc();
643 Constant *EmitArcs = getEmitArcsFunc();
644 Constant *EndFile = getEndFileFunc();
646 for (DebugInfoFinder::iterator CUI = DIF.compile_unit_begin(),
647 CUE = DIF.compile_unit_end(); CUI != CUE; ++CUI) {
648 DICompileUnit compile_unit(*CUI);
649 std::string FilenameGcda = mangleName(compile_unit, "gcda");
650 Builder.CreateCall(StartFile,
651 Builder.CreateGlobalStringPtr(FilenameGcda));
652 for (SmallVector<std::pair<GlobalVariable *, MDNode *>, 8>::iterator
653 I = CountersBySP.begin(), E = CountersBySP.end();
655 DISubprogram SP(I->second);
656 intptr_t ident = reinterpret_cast<intptr_t>(I->second);
657 Builder.CreateCall2(EmitFunction,
658 ConstantInt::get(Type::getInt32Ty(*Ctx), ident),
659 Builder.CreateGlobalStringPtr(SP.getName()));
661 GlobalVariable *GV = I->first;
663 cast<ArrayType>(GV->getType()->getElementType())->getNumElements();
664 Builder.CreateCall2(EmitArcs,
665 ConstantInt::get(Type::getInt32Ty(*Ctx), Arcs),
666 Builder.CreateConstGEP2_64(GV, 0, 0));
668 Builder.CreateCall(EndFile);
670 Builder.CreateRetVoid();
672 InsertProfilingShutdownCall(WriteoutF, M);