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) {
47 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
49 GCOVProfiler(bool EmitNotes, bool EmitData)
50 : ModulePass(ID), EmitNotes(EmitNotes), EmitData(EmitData) {
51 assert((EmitNotes || EmitData) && "GCOVProfiler asked to do nothing?");
52 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
54 virtual const char *getPassName() const {
55 return "GCOV Profiler";
59 bool runOnModule(Module &M);
61 // Create the GCNO files for the Module based on DebugInfo.
62 void emitGCNO(DebugInfoFinder &DIF);
64 // Modify the program to track transitions along edges and call into the
65 // profiling runtime to emit .gcda files when run.
66 bool emitProfileArcs(DebugInfoFinder &DIF);
68 // Get pointers to the functions in the runtime library.
69 Constant *getStartFileFunc();
70 Constant *getIncrementIndirectCounterFunc();
71 Constant *getEmitFunctionFunc();
72 Constant *getEmitArcsFunc();
73 Constant *getEndFileFunc();
75 // Create or retrieve an i32 state value that is used to represent the
76 // pred block number for certain non-trivial edges.
77 GlobalVariable *getEdgeStateValue();
79 // Produce a table of pointers to counters, by predecessor and successor
81 GlobalVariable *buildEdgeLookupTable(Function *F,
82 GlobalVariable *Counter,
83 const UniqueVector<BasicBlock *> &Preds,
84 const UniqueVector<BasicBlock *> &Succs);
86 // Add the function to write out all our counters to the global destructor
88 void insertCounterWriteout(DebugInfoFinder &,
89 SmallVector<std::pair<GlobalVariable *,
92 std::string mangleName(DICompileUnit CU, std::string NewStem);
102 char GCOVProfiler::ID = 0;
103 INITIALIZE_PASS(GCOVProfiler, "insert-gcov-profiling",
104 "Insert instrumentation for GCOV profiling", false, false)
106 ModulePass *llvm::createGCOVProfilerPass(bool EmitNotes, bool EmitData) {
107 return new GCOVProfiler(EmitNotes, EmitData);
110 static DISubprogram findSubprogram(DIScope Scope) {
111 while (!Scope.isSubprogram()) {
112 assert(Scope.isLexicalBlock() &&
113 "Debug location not lexical block or subprogram");
114 Scope = DILexicalBlock(Scope).getContext();
116 return DISubprogram(Scope);
122 static const char *LinesTag;
123 static const char *FunctionTag;
124 static const char *BlockTag;
125 static const char *EdgeTag;
129 void writeBytes(const char *Bytes, int Size) {
130 os->write(Bytes, Size);
133 void write(uint32_t i) {
134 writeBytes(reinterpret_cast<char*>(&i), 4);
137 // Returns the length measured in 4-byte blocks that will be used to
138 // represent this string in a GCOV file
139 unsigned lengthOfGCOVString(StringRef s) {
140 // A GCOV string is a length, followed by a NUL, then between 0 and 3 NULs
141 // padding out to the next 4-byte word. The length is measured in 4-byte
142 // words including padding, not bytes of actual string.
143 return (s.size() + 5) / 4;
146 void writeGCOVString(StringRef s) {
147 uint32_t Len = lengthOfGCOVString(s);
149 writeBytes(s.data(), s.size());
151 // Write 1 to 4 bytes of NUL padding.
152 assert((unsigned)(4 - (s.size() % 4)) > 0);
153 assert((unsigned)(4 - (s.size() % 4)) <= 4);
154 writeBytes("\0\0\0\0", 4 - (s.size() % 4));
159 const char *GCOVRecord::LinesTag = "\0\0\x45\x01";
160 const char *GCOVRecord::FunctionTag = "\0\0\0\1";
161 const char *GCOVRecord::BlockTag = "\0\0\x41\x01";
162 const char *GCOVRecord::EdgeTag = "\0\0\x43\x01";
167 // Constructed only by requesting it from a GCOVBlock, this object stores a
168 // list of line numbers and a single filename, representing lines that belong
170 class GCOVLines : public GCOVRecord {
172 void addLine(uint32_t Line) {
173 Lines.push_back(Line);
177 return lengthOfGCOVString(Filename) + 2 + Lines.size();
181 friend class GCOVBlock;
183 GCOVLines(std::string Filename, raw_ostream *os)
184 : Filename(Filename) {
188 std::string Filename;
189 SmallVector<uint32_t, 32> Lines;
192 // Represent a basic block in GCOV. Each block has a unique number in the
193 // function, number of lines belonging to each block, and a set of edges to
195 class GCOVBlock : public GCOVRecord {
197 GCOVLines &getFile(std::string Filename) {
198 GCOVLines *&Lines = LinesByFile[Filename];
200 Lines = new GCOVLines(Filename, os);
205 void addEdge(GCOVBlock &Successor) {
206 OutEdges.push_back(&Successor);
211 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
212 E = LinesByFile.end(); I != E; ++I) {
213 Len += I->second->length();
216 writeBytes(LinesTag, 4);
219 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
220 E = LinesByFile.end(); I != E; ++I) {
222 writeGCOVString(I->second->Filename);
223 for (int i = 0, e = I->second->Lines.size(); i != e; ++i) {
224 write(I->second->Lines[i]);
232 DeleteContainerSeconds(LinesByFile);
236 friend class GCOVFunction;
238 GCOVBlock(uint32_t Number, raw_ostream *os)
244 StringMap<GCOVLines *> LinesByFile;
245 SmallVector<GCOVBlock *, 4> OutEdges;
248 // A function has a unique identifier, a checksum (we leave as zero) and a
249 // set of blocks and a map of edges between blocks. This is the only GCOV
250 // object users can construct, the blocks and lines will be rooted here.
251 class GCOVFunction : public GCOVRecord {
253 GCOVFunction(DISubprogram SP, raw_ostream *os) {
256 Function *F = SP.getFunction();
258 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
259 Blocks[BB] = new GCOVBlock(i++, os);
261 ReturnBlock = new GCOVBlock(i++, os);
263 writeBytes(FunctionTag, 4);
264 uint32_t BlockLen = 1 + 1 + 1 + lengthOfGCOVString(SP.getName()) +
265 1 + lengthOfGCOVString(SP.getFilename()) + 1;
267 uint32_t Ident = reinterpret_cast<intptr_t>((MDNode*)SP);
269 write(0); // checksum
270 writeGCOVString(SP.getName());
271 writeGCOVString(SP.getFilename());
272 write(SP.getLineNumber());
276 DeleteContainerSeconds(Blocks);
280 GCOVBlock &getBlock(BasicBlock *BB) {
284 GCOVBlock &getReturnBlock() {
289 // Emit count of blocks.
290 writeBytes(BlockTag, 4);
291 write(Blocks.size() + 1);
292 for (int i = 0, e = Blocks.size() + 1; i != e; ++i) {
293 write(0); // No flags on our blocks.
296 // Emit edges between blocks.
297 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
298 E = Blocks.end(); I != E; ++I) {
299 GCOVBlock &Block = *I->second;
300 if (Block.OutEdges.empty()) continue;
302 writeBytes(EdgeTag, 4);
303 write(Block.OutEdges.size() * 2 + 1);
305 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i) {
306 write(Block.OutEdges[i]->Number);
307 write(0); // no flags
311 // Emit lines for each block.
312 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
313 E = Blocks.end(); I != E; ++I) {
314 I->second->writeOut();
319 DenseMap<BasicBlock *, GCOVBlock *> Blocks;
320 GCOVBlock *ReturnBlock;
324 std::string GCOVProfiler::mangleName(DICompileUnit CU, std::string NewStem) {
325 if (NamedMDNode *GCov = M->getNamedMetadata("llvm.gcov")) {
326 for (int i = 0, e = GCov->getNumOperands(); i != e; ++i) {
327 MDNode *N = GCov->getOperand(i);
328 if (N->getNumOperands() != 2) continue;
329 MDString *GCovFile = dyn_cast<MDString>(N->getOperand(0));
330 MDNode *CompileUnit = dyn_cast<MDNode>(N->getOperand(1));
331 if (!GCovFile || !CompileUnit) continue;
332 if (CompileUnit == CU) {
333 SmallString<128> Filename = GCovFile->getString();
334 sys::path::replace_extension(Filename, NewStem);
335 return Filename.str();
340 SmallString<128> Filename = CU.getFilename();
341 sys::path::replace_extension(Filename, NewStem);
342 return sys::path::filename(Filename.str());
345 bool GCOVProfiler::runOnModule(Module &M) {
347 Ctx = &M.getContext();
350 DIF.processModule(M);
352 if (EmitNotes) emitGCNO(DIF);
353 if (EmitData) return emitProfileArcs(DIF);
357 void GCOVProfiler::emitGCNO(DebugInfoFinder &DIF) {
358 DenseMap<const MDNode *, raw_fd_ostream *> GcnoFiles;
359 for (DebugInfoFinder::iterator I = DIF.compile_unit_begin(),
360 E = DIF.compile_unit_end(); I != E; ++I) {
361 // Each compile unit gets its own .gcno file. This means that whether we run
362 // this pass over the original .o's as they're produced, or run it after
363 // LTO, we'll generate the same .gcno files.
365 DICompileUnit CU(*I);
366 raw_fd_ostream *&out = GcnoFiles[CU];
367 std::string ErrorInfo;
368 out = new raw_fd_ostream(mangleName(CU, "gcno").c_str(), ErrorInfo,
369 raw_fd_ostream::F_Binary);
370 out->write("oncg*404MVLL", 12);
373 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
374 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
375 DISubprogram SP(*SPI);
376 raw_fd_ostream *&os = GcnoFiles[SP.getCompileUnit()];
378 Function *F = SP.getFunction();
380 GCOVFunction Func(SP, os);
382 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
383 GCOVBlock &Block = Func.getBlock(BB);
384 TerminatorInst *TI = BB->getTerminator();
385 if (int successors = TI->getNumSuccessors()) {
386 for (int i = 0; i != successors; ++i) {
387 Block.addEdge(Func.getBlock(TI->getSuccessor(i)));
389 } else if (isa<ReturnInst>(TI)) {
390 Block.addEdge(Func.getReturnBlock());
394 for (BasicBlock::iterator I = BB->begin(), IE = BB->end(); I != IE; ++I) {
395 const DebugLoc &Loc = I->getDebugLoc();
396 if (Loc.isUnknown()) continue;
397 if (Line == Loc.getLine()) continue;
398 Line = Loc.getLine();
399 if (SP != findSubprogram(DIScope(Loc.getScope(*Ctx)))) continue;
401 GCOVLines &Lines = Block.getFile(SP.getFilename());
402 Lines.addLine(Loc.getLine());
408 for (DenseMap<const MDNode *, raw_fd_ostream *>::iterator
409 I = GcnoFiles.begin(), E = GcnoFiles.end(); I != E; ++I) {
410 raw_fd_ostream *&out = I->second;
411 out->write("\0\0\0\0\0\0\0\0", 8); // EOF
417 bool GCOVProfiler::emitProfileArcs(DebugInfoFinder &DIF) {
418 if (DIF.subprogram_begin() == DIF.subprogram_end())
421 SmallVector<std::pair<GlobalVariable *, uint32_t>, 8> CountersByIdent;
422 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
423 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
424 DISubprogram SP(*SPI);
425 Function *F = SP.getFunction();
429 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
430 TerminatorInst *TI = BB->getTerminator();
431 if (isa<ReturnInst>(TI))
434 Edges += TI->getNumSuccessors();
437 const ArrayType *CounterTy =
438 ArrayType::get(Type::getInt64Ty(*Ctx), Edges);
439 GlobalVariable *Counters =
440 new GlobalVariable(*M, CounterTy, false,
441 GlobalValue::InternalLinkage,
442 Constant::getNullValue(CounterTy),
443 "__llvm_gcov_ctr", 0, false, 0);
444 CountersByIdent.push_back(
445 std::make_pair(Counters, reinterpret_cast<intptr_t>((MDNode*)SP)));
447 UniqueVector<BasicBlock *> ComplexEdgePreds;
448 UniqueVector<BasicBlock *> ComplexEdgeSuccs;
451 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
452 TerminatorInst *TI = BB->getTerminator();
453 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
455 IRBuilder<> Builder(TI);
457 if (Successors == 1) {
458 Value *Counter = Builder.CreateConstInBoundsGEP2_64(Counters, 0,
460 Value *Count = Builder.CreateLoad(Counter);
461 Count = Builder.CreateAdd(Count,
462 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
463 Builder.CreateStore(Count, Counter);
464 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
465 Value *Sel = Builder.CreateSelect(
467 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge),
468 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge + 1));
469 SmallVector<Value *, 2> Idx;
470 Idx.push_back(Constant::getNullValue(Type::getInt64Ty(*Ctx)));
472 Value *Counter = Builder.CreateInBoundsGEP(Counters,
473 Idx.begin(), Idx.end());
474 Value *Count = Builder.CreateLoad(Counter);
475 Count = Builder.CreateAdd(Count,
476 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
477 Builder.CreateStore(Count, Counter);
479 ComplexEdgePreds.insert(BB);
480 for (int i = 0; i != Successors; ++i)
481 ComplexEdgeSuccs.insert(TI->getSuccessor(i));
487 if (!ComplexEdgePreds.empty()) {
488 GlobalVariable *EdgeTable =
489 buildEdgeLookupTable(F, Counters,
490 ComplexEdgePreds, ComplexEdgeSuccs);
491 GlobalVariable *EdgeState = getEdgeStateValue();
493 const Type *Int32Ty = Type::getInt32Ty(*Ctx);
494 for (int i = 0, e = ComplexEdgePreds.size(); i != e; ++i) {
495 IRBuilder<> Builder(ComplexEdgePreds[i+1]->getTerminator());
496 Builder.CreateStore(ConstantInt::get(Int32Ty, i), EdgeState);
498 for (int i = 0, e = ComplexEdgeSuccs.size(); i != e; ++i) {
499 // call runtime to perform increment
500 IRBuilder<> Builder(ComplexEdgeSuccs[i+1]->getFirstNonPHI());
501 Value *CounterPtrArray =
502 Builder.CreateConstInBoundsGEP2_64(EdgeTable, 0,
503 i * ComplexEdgePreds.size());
504 Builder.CreateCall2(getIncrementIndirectCounterFunc(),
505 EdgeState, CounterPtrArray);
506 // clear the predecessor number
507 Builder.CreateStore(ConstantInt::get(Int32Ty, 0xffffffff), EdgeState);
512 insertCounterWriteout(DIF, CountersByIdent);
517 // All edges with successors that aren't branches are "complex", because it
518 // requires complex logic to pick which counter to update.
519 GlobalVariable *GCOVProfiler::buildEdgeLookupTable(
521 GlobalVariable *Counters,
522 const UniqueVector<BasicBlock *> &Preds,
523 const UniqueVector<BasicBlock *> &Succs) {
524 // TODO: support invoke, threads. We rely on the fact that nothing can modify
525 // the whole-Module pred edge# between the time we set it and the time we next
526 // read it. Threads and invoke make this untrue.
528 // emit [(succs * preds) x i64*], logically [succ x [pred x i64*]].
529 const Type *Int64PtrTy = Type::getInt64PtrTy(*Ctx);
530 const ArrayType *EdgeTableTy = ArrayType::get(
531 Int64PtrTy, Succs.size() * Preds.size());
533 Constant **EdgeTable = new Constant*[Succs.size() * Preds.size()];
534 Constant *NullValue = Constant::getNullValue(Int64PtrTy);
535 for (int i = 0, ie = Succs.size() * Preds.size(); i != ie; ++i)
536 EdgeTable[i] = NullValue;
539 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
540 TerminatorInst *TI = BB->getTerminator();
541 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
542 if (Successors > 1 && !isa<BranchInst>(TI) && !isa<ReturnInst>(TI)) {
543 for (int i = 0; i != Successors; ++i) {
544 BasicBlock *Succ = TI->getSuccessor(i);
545 IRBuilder<> builder(Succ);
546 Value *Counter = builder.CreateConstInBoundsGEP2_64(Counters, 0,
548 EdgeTable[((Succs.idFor(Succ)-1) * Preds.size()) +
549 (Preds.idFor(BB)-1)] = cast<Constant>(Counter);
555 GlobalVariable *EdgeTableGV =
557 *M, EdgeTableTy, true, GlobalValue::InternalLinkage,
558 ConstantArray::get(EdgeTableTy,
559 &EdgeTable[0], Succs.size() * Preds.size()),
560 "__llvm_gcda_edge_table");
561 EdgeTableGV->setUnnamedAddr(true);
565 Constant *GCOVProfiler::getStartFileFunc() {
566 const Type *Args[] = { Type::getInt8PtrTy(*Ctx) };
567 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
569 return M->getOrInsertFunction("llvm_gcda_start_file", FTy);
572 Constant *GCOVProfiler::getIncrementIndirectCounterFunc() {
573 const Type *Args[] = {
574 Type::getInt32PtrTy(*Ctx), // uint32_t *predecessor
575 Type::getInt64PtrTy(*Ctx)->getPointerTo(), // uint64_t **state_table_row
577 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
579 return M->getOrInsertFunction("llvm_gcda_increment_indirect_counter", FTy);
582 Constant *GCOVProfiler::getEmitFunctionFunc() {
583 const Type *Args[] = { Type::getInt32Ty(*Ctx) };
584 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
586 return M->getOrInsertFunction("llvm_gcda_emit_function", FTy);
589 Constant *GCOVProfiler::getEmitArcsFunc() {
590 const Type *Args[] = {
591 Type::getInt32Ty(*Ctx), // uint32_t num_counters
592 Type::getInt64PtrTy(*Ctx), // uint64_t *counters
594 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
596 return M->getOrInsertFunction("llvm_gcda_emit_arcs", FTy);
599 Constant *GCOVProfiler::getEndFileFunc() {
600 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
601 return M->getOrInsertFunction("llvm_gcda_end_file", FTy);
604 GlobalVariable *GCOVProfiler::getEdgeStateValue() {
605 GlobalVariable *GV = M->getGlobalVariable("__llvm_gcov_global_state_pred");
607 GV = new GlobalVariable(*M, Type::getInt32Ty(*Ctx), false,
608 GlobalValue::InternalLinkage,
609 ConstantInt::get(Type::getInt32Ty(*Ctx),
611 "__llvm_gcov_global_state_pred");
612 GV->setUnnamedAddr(true);
617 void GCOVProfiler::insertCounterWriteout(
618 DebugInfoFinder &DIF,
619 SmallVector<std::pair<GlobalVariable *, uint32_t>, 8> &CountersByIdent) {
620 const FunctionType *WriteoutFTy =
621 FunctionType::get(Type::getVoidTy(*Ctx), false);
622 Function *WriteoutF = Function::Create(WriteoutFTy,
623 GlobalValue::InternalLinkage,
624 "__llvm_gcov_writeout", M);
625 WriteoutF->setUnnamedAddr(true);
626 BasicBlock *BB = BasicBlock::Create(*Ctx, "", WriteoutF);
627 IRBuilder<> Builder(BB);
629 Constant *StartFile = getStartFileFunc();
630 Constant *EmitFunction = getEmitFunctionFunc();
631 Constant *EmitArcs = getEmitArcsFunc();
632 Constant *EndFile = getEndFileFunc();
634 for (DebugInfoFinder::iterator CUI = DIF.compile_unit_begin(),
635 CUE = DIF.compile_unit_end(); CUI != CUE; ++CUI) {
636 DICompileUnit compile_unit(*CUI);
637 std::string FilenameGcda = mangleName(compile_unit, "gcda");
638 Builder.CreateCall(StartFile,
639 Builder.CreateGlobalStringPtr(FilenameGcda));
640 for (SmallVector<std::pair<GlobalVariable *, uint32_t>, 8>::iterator
641 I = CountersByIdent.begin(), E = CountersByIdent.end();
643 Builder.CreateCall(EmitFunction, ConstantInt::get(Type::getInt32Ty(*Ctx),
645 GlobalVariable *GV = I->first;
647 cast<ArrayType>(GV->getType()->getElementType())->getNumElements();
648 Builder.CreateCall2(EmitArcs,
649 ConstantInt::get(Type::getInt32Ty(*Ctx), Arcs),
650 Builder.CreateConstGEP2_64(GV, 0, 0));
652 Builder.CreateCall(EndFile);
654 Builder.CreateRetVoid();
656 InsertProfilingShutdownCall(WriteoutF, M);