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() / 4) + 1;
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 + 1 + lengthOfGCOVString(SP.getName()) +
265 1 + lengthOfGCOVString(SP.getFilename()) + 1;
267 uint32_t Ident = reinterpret_cast<intptr_t>((MDNode*)SP);
269 write(0); // checksum #1
270 write(0); // checksum #2
271 writeGCOVString(SP.getName());
272 writeGCOVString(SP.getFilename());
273 write(SP.getLineNumber());
277 DeleteContainerSeconds(Blocks);
281 GCOVBlock &getBlock(BasicBlock *BB) {
285 GCOVBlock &getReturnBlock() {
290 // Emit count of blocks.
291 writeBytes(BlockTag, 4);
292 write(Blocks.size() + 1);
293 for (int i = 0, e = Blocks.size() + 1; i != e; ++i) {
294 write(0); // No flags on our blocks.
297 // Emit edges between blocks.
298 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
299 E = Blocks.end(); I != E; ++I) {
300 GCOVBlock &Block = *I->second;
301 if (Block.OutEdges.empty()) continue;
303 writeBytes(EdgeTag, 4);
304 write(Block.OutEdges.size() * 2 + 1);
306 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i) {
307 write(Block.OutEdges[i]->Number);
308 write(0); // no flags
312 // Emit lines for each block.
313 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
314 E = Blocks.end(); I != E; ++I) {
315 I->second->writeOut();
320 DenseMap<BasicBlock *, GCOVBlock *> Blocks;
321 GCOVBlock *ReturnBlock;
325 std::string GCOVProfiler::mangleName(DICompileUnit CU, std::string NewStem) {
326 if (NamedMDNode *GCov = M->getNamedMetadata("llvm.gcov")) {
327 for (int i = 0, e = GCov->getNumOperands(); i != e; ++i) {
328 MDNode *N = GCov->getOperand(i);
329 if (N->getNumOperands() != 2) continue;
330 MDString *GCovFile = dyn_cast<MDString>(N->getOperand(0));
331 MDNode *CompileUnit = dyn_cast<MDNode>(N->getOperand(1));
332 if (!GCovFile || !CompileUnit) continue;
333 if (CompileUnit == CU) {
334 SmallString<128> Filename = GCovFile->getString();
335 sys::path::replace_extension(Filename, NewStem);
336 return Filename.str();
341 SmallString<128> Filename = CU.getFilename();
342 sys::path::replace_extension(Filename, NewStem);
343 return sys::path::filename(Filename.str());
346 bool GCOVProfiler::runOnModule(Module &M) {
348 Ctx = &M.getContext();
351 DIF.processModule(M);
353 if (EmitNotes) emitGCNO(DIF);
354 if (EmitData) return emitProfileArcs(DIF);
358 void GCOVProfiler::emitGCNO(DebugInfoFinder &DIF) {
359 DenseMap<const MDNode *, raw_fd_ostream *> GcnoFiles;
360 for (DebugInfoFinder::iterator I = DIF.compile_unit_begin(),
361 E = DIF.compile_unit_end(); I != E; ++I) {
362 // Each compile unit gets its own .gcno file. This means that whether we run
363 // this pass over the original .o's as they're produced, or run it after
364 // LTO, we'll generate the same .gcno files.
366 DICompileUnit CU(*I);
367 raw_fd_ostream *&out = GcnoFiles[CU];
368 std::string ErrorInfo;
369 out = new raw_fd_ostream(mangleName(CU, "gcno").c_str(), ErrorInfo,
370 raw_fd_ostream::F_Binary);
371 out->write("oncg*404MVLL", 12);
374 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
375 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
376 DISubprogram SP(*SPI);
377 raw_fd_ostream *&os = GcnoFiles[SP.getCompileUnit()];
379 Function *F = SP.getFunction();
381 GCOVFunction Func(SP, os);
383 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
384 GCOVBlock &Block = Func.getBlock(BB);
385 TerminatorInst *TI = BB->getTerminator();
386 if (int successors = TI->getNumSuccessors()) {
387 for (int i = 0; i != successors; ++i) {
388 Block.addEdge(Func.getBlock(TI->getSuccessor(i)));
390 } else if (isa<ReturnInst>(TI)) {
391 Block.addEdge(Func.getReturnBlock());
395 for (BasicBlock::iterator I = BB->begin(), IE = BB->end(); I != IE; ++I) {
396 const DebugLoc &Loc = I->getDebugLoc();
397 if (Loc.isUnknown()) continue;
398 if (Line == Loc.getLine()) continue;
399 Line = Loc.getLine();
400 if (SP != findSubprogram(DIScope(Loc.getScope(*Ctx)))) continue;
402 GCOVLines &Lines = Block.getFile(SP.getFilename());
403 Lines.addLine(Loc.getLine());
409 for (DenseMap<const MDNode *, raw_fd_ostream *>::iterator
410 I = GcnoFiles.begin(), E = GcnoFiles.end(); I != E; ++I) {
411 raw_fd_ostream *&out = I->second;
412 out->write("\0\0\0\0\0\0\0\0", 8); // EOF
418 bool GCOVProfiler::emitProfileArcs(DebugInfoFinder &DIF) {
419 if (DIF.subprogram_begin() == DIF.subprogram_end())
422 SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> CountersBySP;
423 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
424 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
425 DISubprogram SP(*SPI);
426 Function *F = SP.getFunction();
430 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
431 TerminatorInst *TI = BB->getTerminator();
432 if (isa<ReturnInst>(TI))
435 Edges += TI->getNumSuccessors();
438 const ArrayType *CounterTy =
439 ArrayType::get(Type::getInt64Ty(*Ctx), Edges);
440 GlobalVariable *Counters =
441 new GlobalVariable(*M, CounterTy, false,
442 GlobalValue::InternalLinkage,
443 Constant::getNullValue(CounterTy),
444 "__llvm_gcov_ctr", 0, false, 0);
445 CountersBySP.push_back(std::make_pair(Counters, (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, CountersBySP);
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[2] = {
584 Type::getInt32Ty(*Ctx), // uint32_t ident
585 Type::getInt8PtrTy(*Ctx), // const char *function_name
587 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
589 return M->getOrInsertFunction("llvm_gcda_emit_function", FTy);
592 Constant *GCOVProfiler::getEmitArcsFunc() {
593 const Type *Args[] = {
594 Type::getInt32Ty(*Ctx), // uint32_t num_counters
595 Type::getInt64PtrTy(*Ctx), // uint64_t *counters
597 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
599 return M->getOrInsertFunction("llvm_gcda_emit_arcs", FTy);
602 Constant *GCOVProfiler::getEndFileFunc() {
603 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
604 return M->getOrInsertFunction("llvm_gcda_end_file", FTy);
607 GlobalVariable *GCOVProfiler::getEdgeStateValue() {
608 GlobalVariable *GV = M->getGlobalVariable("__llvm_gcov_global_state_pred");
610 GV = new GlobalVariable(*M, Type::getInt32Ty(*Ctx), false,
611 GlobalValue::InternalLinkage,
612 ConstantInt::get(Type::getInt32Ty(*Ctx),
614 "__llvm_gcov_global_state_pred");
615 GV->setUnnamedAddr(true);
620 void GCOVProfiler::insertCounterWriteout(
621 DebugInfoFinder &DIF,
622 SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> &CountersBySP) {
623 const FunctionType *WriteoutFTy =
624 FunctionType::get(Type::getVoidTy(*Ctx), false);
625 Function *WriteoutF = Function::Create(WriteoutFTy,
626 GlobalValue::InternalLinkage,
627 "__llvm_gcov_writeout", M);
628 WriteoutF->setUnnamedAddr(true);
629 BasicBlock *BB = BasicBlock::Create(*Ctx, "", WriteoutF);
630 IRBuilder<> Builder(BB);
632 Constant *StartFile = getStartFileFunc();
633 Constant *EmitFunction = getEmitFunctionFunc();
634 Constant *EmitArcs = getEmitArcsFunc();
635 Constant *EndFile = getEndFileFunc();
637 for (DebugInfoFinder::iterator CUI = DIF.compile_unit_begin(),
638 CUE = DIF.compile_unit_end(); CUI != CUE; ++CUI) {
639 DICompileUnit compile_unit(*CUI);
640 std::string FilenameGcda = mangleName(compile_unit, "gcda");
641 Builder.CreateCall(StartFile,
642 Builder.CreateGlobalStringPtr(FilenameGcda));
643 for (SmallVector<std::pair<GlobalVariable *, MDNode *>, 8>::iterator
644 I = CountersBySP.begin(), E = CountersBySP.end();
646 DISubprogram SP(I->second);
647 intptr_t ident = reinterpret_cast<intptr_t>(I->second);
648 Builder.CreateCall2(EmitFunction,
649 ConstantInt::get(Type::getInt32Ty(*Ctx), ident),
650 Builder.CreateGlobalStringPtr(SP.getName()));
652 GlobalVariable *GV = I->first;
654 cast<ArrayType>(GV->getType()->getElementType())->getNumElements();
655 Builder.CreateCall2(EmitArcs,
656 ConstantInt::get(Type::getInt32Ty(*Ctx), Arcs),
657 Builder.CreateConstGEP2_64(GV, 0, 0));
659 Builder.CreateCall(EndFile);
661 Builder.CreateRetVoid();
663 InsertProfilingShutdownCall(WriteoutF, M);