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 {
43 bool runOnModule(Module &M);
47 : ModulePass(ID), EmitNotes(true), EmitData(true) {
48 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
50 GCOVProfiler(bool EmitNotes, bool EmitData)
51 : ModulePass(ID), EmitNotes(EmitNotes), EmitData(EmitData) {
52 assert((EmitNotes || EmitData) && "GCOVProfiler asked to do nothing?");
53 initializeGCOVProfilerPass(*PassRegistry::getPassRegistry());
55 virtual const char *getPassName() const {
56 return "GCOV Profiler";
60 // Create the GCNO files for the Module based on DebugInfo.
61 void emitGCNO(DebugInfoFinder &DIF);
63 // Modify the program to track transitions along edges and call into the
64 // profiling runtime to emit .gcda files when run.
65 bool emitProfileArcs(DebugInfoFinder &DIF);
67 // Get pointers to the functions in the runtime library.
68 Constant *getStartFileFunc();
69 Constant *getIncrementIndirectCounterFunc();
70 Constant *getEmitFunctionFunc();
71 Constant *getEmitArcsFunc();
72 Constant *getEndFileFunc();
74 // Create or retrieve an i32 state value that is used to represent the
75 // pred block number for certain non-trivial edges.
76 GlobalVariable *getEdgeStateValue();
78 // Produce a table of pointers to counters, by predecessor and successor
80 GlobalVariable *buildEdgeLookupTable(Function *F,
81 GlobalVariable *Counter,
82 const UniqueVector<BasicBlock *> &Preds,
83 const UniqueVector<BasicBlock *> &Succs);
85 // Add the function to write out all our counters to the global destructor
87 void insertCounterWriteout(DebugInfoFinder &,
88 SmallVector<std::pair<GlobalVariable *,
99 char GCOVProfiler::ID = 0;
100 INITIALIZE_PASS(GCOVProfiler, "insert-gcov-profiling",
101 "Insert instrumentation for GCOV profiling", false, false)
103 ModulePass *llvm::createGCOVProfilerPass(bool EmitNotes, bool EmitData) {
104 return new GCOVProfiler(EmitNotes, EmitData);
107 static DISubprogram findSubprogram(DIScope Scope) {
108 while (!Scope.isSubprogram()) {
109 assert(Scope.isLexicalBlock() &&
110 "Debug location not lexical block or subprogram");
111 Scope = DILexicalBlock(Scope).getContext();
113 return DISubprogram(Scope);
119 static const char *LinesTag;
120 static const char *FunctionTag;
121 static const char *BlockTag;
122 static const char *EdgeTag;
126 void writeBytes(const char *Bytes, int Size) {
127 os->write(Bytes, Size);
130 void write(uint32_t i) {
131 writeBytes(reinterpret_cast<char*>(&i), 4);
134 // Returns the length measured in 4-byte blocks that will be used to
135 // represent this string in a GCOV file
136 unsigned lengthOfGCOVString(StringRef s) {
137 // A GCOV string is a length, followed by a NUL, then between 0 and 3 NULs
138 // padding out to the next 4-byte word. The length is measured in 4-byte
139 // words including padding, not bytes of actual string.
140 return (s.size() + 5) / 4;
143 void writeGCOVString(StringRef s) {
144 uint32_t Len = lengthOfGCOVString(s);
146 writeBytes(s.data(), s.size());
148 // Write 1 to 4 bytes of NUL padding.
149 assert((unsigned)(5 - ((s.size() + 1) % 4)) > 0);
150 assert((unsigned)(5 - ((s.size() + 1) % 4)) <= 4);
151 writeBytes("\0\0\0\0", 5 - ((s.size() + 1) % 4));
156 const char *GCOVRecord::LinesTag = "\0\0\x45\x01";
157 const char *GCOVRecord::FunctionTag = "\0\0\0\1";
158 const char *GCOVRecord::BlockTag = "\0\0\x41\x01";
159 const char *GCOVRecord::EdgeTag = "\0\0\x43\x01";
164 // Constructed only by requesting it from a GCOVBlock, this object stores a
165 // list of line numbers and a single filename, representing lines that belong
167 class GCOVLines : public GCOVRecord {
169 void addLine(uint32_t Line) {
170 Lines.push_back(Line);
174 return lengthOfGCOVString(Filename) + 2 + Lines.size();
178 friend class GCOVBlock;
180 GCOVLines(std::string Filename, raw_ostream *os)
181 : Filename(Filename) {
185 std::string Filename;
186 SmallVector<uint32_t, 32> Lines;
189 // Represent a basic block in GCOV. Each block has a unique number in the
190 // function, number of lines belonging to each block, and a set of edges to
192 class GCOVBlock : public GCOVRecord {
194 GCOVLines &getFile(std::string Filename) {
195 GCOVLines *&Lines = LinesByFile[Filename];
197 Lines = new GCOVLines(Filename, os);
202 void addEdge(GCOVBlock &Successor) {
203 OutEdges.push_back(&Successor);
208 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
209 E = LinesByFile.end(); I != E; ++I) {
210 Len += I->second->length();
213 writeBytes(LinesTag, 4);
216 for (StringMap<GCOVLines *>::iterator I = LinesByFile.begin(),
217 E = LinesByFile.end(); I != E; ++I) {
219 writeGCOVString(I->second->Filename);
220 for (int i = 0, e = I->second->Lines.size(); i != e; ++i) {
221 write(I->second->Lines[i]);
229 DeleteContainerSeconds(LinesByFile);
233 friend class GCOVFunction;
235 GCOVBlock(uint32_t Number, raw_ostream *os)
241 StringMap<GCOVLines *> LinesByFile;
242 SmallVector<GCOVBlock *, 4> OutEdges;
245 // A function has a unique identifier, a checksum (we leave as zero) and a
246 // set of blocks and a map of edges between blocks. This is the only GCOV
247 // object users can construct, the blocks and lines will be rooted here.
248 class GCOVFunction : public GCOVRecord {
250 GCOVFunction(DISubprogram SP, raw_ostream *os) {
253 Function *F = SP.getFunction();
255 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
256 Blocks[BB] = new GCOVBlock(i++, os);
258 ReturnBlock = new GCOVBlock(i++, os);
260 writeBytes(FunctionTag, 4);
261 uint32_t BlockLen = 1 + 1 + 1 + lengthOfGCOVString(SP.getName()) +
262 1 + lengthOfGCOVString(SP.getFilename()) + 1;
264 uint32_t Ident = reinterpret_cast<intptr_t>((MDNode*)SP);
266 write(0); // checksum
267 writeGCOVString(SP.getName());
268 writeGCOVString(SP.getFilename());
269 write(SP.getLineNumber());
273 DeleteContainerSeconds(Blocks);
277 GCOVBlock &getBlock(BasicBlock *BB) {
281 GCOVBlock &getReturnBlock() {
286 // Emit count of blocks.
287 writeBytes(BlockTag, 4);
288 write(Blocks.size() + 1);
289 for (int i = 0, e = Blocks.size() + 1; i != e; ++i) {
290 write(0); // No flags on our blocks.
293 // Emit edges between blocks.
294 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
295 E = Blocks.end(); I != E; ++I) {
296 GCOVBlock &Block = *I->second;
297 if (Block.OutEdges.empty()) continue;
299 writeBytes(EdgeTag, 4);
300 write(Block.OutEdges.size() * 2 + 1);
302 for (int i = 0, e = Block.OutEdges.size(); i != e; ++i) {
303 write(Block.OutEdges[i]->Number);
304 write(0); // no flags
308 // Emit lines for each block.
309 for (DenseMap<BasicBlock *, GCOVBlock *>::iterator I = Blocks.begin(),
310 E = Blocks.end(); I != E; ++I) {
311 I->second->writeOut();
316 DenseMap<BasicBlock *, GCOVBlock *> Blocks;
317 GCOVBlock *ReturnBlock;
321 // Replace the stem of a file, or add one if missing.
322 static std::string replaceStem(std::string OrigFilename, std::string NewStem) {
323 return (sys::path::stem(OrigFilename) + "." + NewStem).str();
326 bool GCOVProfiler::runOnModule(Module &M) {
328 Ctx = &M.getContext();
331 DIF.processModule(M);
333 if (EmitNotes) emitGCNO(DIF);
334 if (EmitData) return emitProfileArcs(DIF);
338 void GCOVProfiler::emitGCNO(DebugInfoFinder &DIF) {
339 DenseMap<const MDNode *, raw_fd_ostream *> GcnoFiles;
340 for (DebugInfoFinder::iterator I = DIF.compile_unit_begin(),
341 E = DIF.compile_unit_end(); I != E; ++I) {
342 // Each compile unit gets its own .gcno file. This means that whether we run
343 // this pass over the original .o's as they're produced, or run it after
344 // LTO, we'll generate the same .gcno files.
346 DICompileUnit CU(*I);
347 raw_fd_ostream *&out = GcnoFiles[CU];
348 std::string ErrorInfo;
349 out = new raw_fd_ostream(replaceStem(CU.getFilename(), "gcno").c_str(),
350 ErrorInfo, raw_fd_ostream::F_Binary);
351 out->write("oncg*404MVLL", 12);
354 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
355 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
356 DISubprogram SP(*SPI);
357 raw_fd_ostream *&os = GcnoFiles[SP.getCompileUnit()];
359 GCOVFunction Func(SP, os);
360 Function *F = SP.getFunction();
361 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
362 GCOVBlock &Block = Func.getBlock(BB);
363 TerminatorInst *TI = BB->getTerminator();
364 if (int successors = TI->getNumSuccessors()) {
365 for (int i = 0; i != successors; ++i) {
366 Block.addEdge(Func.getBlock(TI->getSuccessor(i)));
368 } else if (isa<ReturnInst>(TI)) {
369 Block.addEdge(Func.getReturnBlock());
373 for (BasicBlock::iterator I = BB->begin(), IE = BB->end(); I != IE; ++I) {
374 const DebugLoc &Loc = I->getDebugLoc();
375 if (Loc.isUnknown()) continue;
376 if (Line == Loc.getLine()) continue;
377 Line = Loc.getLine();
378 if (SP != findSubprogram(DIScope(Loc.getScope(*Ctx)))) continue;
380 GCOVLines &Lines = Block.getFile(SP.getFilename());
381 Lines.addLine(Loc.getLine());
387 for (DenseMap<const MDNode *, raw_fd_ostream *>::iterator
388 I = GcnoFiles.begin(), E = GcnoFiles.end(); I != E; ++I) {
389 raw_fd_ostream *&out = I->second;
390 out->write("\0\0\0\0\0\0\0\0", 8); // EOF
396 bool GCOVProfiler::emitProfileArcs(DebugInfoFinder &DIF) {
397 if (DIF.subprogram_begin() == DIF.subprogram_end())
400 SmallVector<std::pair<GlobalVariable *, uint32_t>, 8> CountersByIdent;
401 for (DebugInfoFinder::iterator SPI = DIF.subprogram_begin(),
402 SPE = DIF.subprogram_end(); SPI != SPE; ++SPI) {
403 DISubprogram SP(*SPI);
404 Function *F = SP.getFunction();
407 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
408 TerminatorInst *TI = BB->getTerminator();
409 if (isa<ReturnInst>(TI))
412 Edges += TI->getNumSuccessors();
415 const ArrayType *CounterTy =
416 ArrayType::get(Type::getInt64Ty(*Ctx), Edges);
417 GlobalVariable *Counters =
418 new GlobalVariable(*M, CounterTy, false,
419 GlobalValue::InternalLinkage,
420 Constant::getNullValue(CounterTy),
421 "__llvm_gcov_ctr", 0, false, 0);
422 CountersByIdent.push_back(
423 std::make_pair(Counters, reinterpret_cast<intptr_t>((MDNode*)SP)));
425 UniqueVector<BasicBlock *> ComplexEdgePreds;
426 UniqueVector<BasicBlock *> ComplexEdgeSuccs;
429 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
430 TerminatorInst *TI = BB->getTerminator();
431 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
433 IRBuilder<> Builder(TI);
435 if (Successors == 1) {
436 Value *Counter = Builder.CreateConstInBoundsGEP2_64(Counters, 0,
438 Value *Count = Builder.CreateLoad(Counter);
439 Count = Builder.CreateAdd(Count,
440 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
441 Builder.CreateStore(Count, Counter);
442 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
443 Value *Sel = Builder.CreateSelect(
445 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge),
446 ConstantInt::get(Type::getInt64Ty(*Ctx), Edge + 1));
447 SmallVector<Value *, 2> Idx;
448 Idx.push_back(Constant::getNullValue(Type::getInt64Ty(*Ctx)));
450 Value *Counter = Builder.CreateInBoundsGEP(Counters,
451 Idx.begin(), Idx.end());
452 Value *Count = Builder.CreateLoad(Counter);
453 Count = Builder.CreateAdd(Count,
454 ConstantInt::get(Type::getInt64Ty(*Ctx),1));
455 Builder.CreateStore(Count, Counter);
457 ComplexEdgePreds.insert(BB);
458 for (int i = 0; i != Successors; ++i)
459 ComplexEdgeSuccs.insert(TI->getSuccessor(i));
465 if (!ComplexEdgePreds.empty()) {
466 GlobalVariable *EdgeTable =
467 buildEdgeLookupTable(F, Counters,
468 ComplexEdgePreds, ComplexEdgeSuccs);
469 GlobalVariable *EdgeState = getEdgeStateValue();
471 const Type *Int32Ty = Type::getInt32Ty(*Ctx);
472 for (int i = 0, e = ComplexEdgePreds.size(); i != e; ++i) {
473 IRBuilder<> Builder(ComplexEdgePreds[i+1]->getTerminator());
474 Builder.CreateStore(ConstantInt::get(Int32Ty, i+1), EdgeState);
476 for (int i = 0, e = ComplexEdgeSuccs.size(); i != e; ++i) {
477 // call runtime to perform increment
478 IRBuilder<> Builder(ComplexEdgeSuccs[i+1]->getFirstNonPHI());
479 Value *CounterPtrArray =
480 Builder.CreateConstInBoundsGEP2_64(EdgeTable, 0,
481 i * ComplexEdgePreds.size());
482 Builder.CreateCall2(getIncrementIndirectCounterFunc(),
483 EdgeState, CounterPtrArray);
484 // clear the predecessor number
485 Builder.CreateStore(ConstantInt::get(Int32Ty, 0xffffffff), EdgeState);
490 insertCounterWriteout(DIF, CountersByIdent);
495 // All edges with successors that aren't branches are "complex", because it
496 // requires complex logic to pick which counter to update.
497 GlobalVariable *GCOVProfiler::buildEdgeLookupTable(
499 GlobalVariable *Counters,
500 const UniqueVector<BasicBlock *> &Preds,
501 const UniqueVector<BasicBlock *> &Succs) {
502 // TODO: support invoke, threads. We rely on the fact that nothing can modify
503 // the whole-Module pred edge# between the time we set it and the time we next
504 // read it. Threads and invoke make this untrue.
506 // emit [(succs * preds) x i64*], logically [succ x [pred x i64*]].
507 const Type *Int64PtrTy = Type::getInt64PtrTy(*Ctx);
508 const ArrayType *EdgeTableTy = ArrayType::get(
509 Int64PtrTy, Succs.size() * Preds.size());
511 Constant **EdgeTable = new Constant*[Succs.size() * Preds.size()];
512 Constant *NullValue = Constant::getNullValue(Int64PtrTy);
513 for (int i = 0, ie = Succs.size() * Preds.size(); i != ie; ++i)
514 EdgeTable[i] = NullValue;
517 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
518 TerminatorInst *TI = BB->getTerminator();
519 int Successors = isa<ReturnInst>(TI) ? 1 : TI->getNumSuccessors();
520 if (Successors && !isa<BranchInst>(TI) && !isa<ReturnInst>(TI)) {
521 for (int i = 0; i != Successors; ++i) {
522 BasicBlock *Succ = TI->getSuccessor(i);
523 IRBuilder<> builder(Succ);
524 Value *Counter = builder.CreateConstInBoundsGEP2_64(Counters, 0,
526 EdgeTable[((Succs.idFor(Succ)-1) * Preds.size()) +
527 (Preds.idFor(BB)-1)] = cast<Constant>(Counter);
533 GlobalVariable *EdgeTableGV =
535 *M, EdgeTableTy, true, GlobalValue::InternalLinkage,
536 ConstantArray::get(EdgeTableTy,
537 &EdgeTable[0], Succs.size() * Preds.size()),
538 "__llvm_gcda_edge_table");
539 EdgeTableGV->setUnnamedAddr(true);
543 Constant *GCOVProfiler::getStartFileFunc() {
544 const Type *Args[] = { Type::getInt8PtrTy(*Ctx) };
545 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
547 return M->getOrInsertFunction("llvm_gcda_start_file", FTy);
550 Constant *GCOVProfiler::getIncrementIndirectCounterFunc() {
551 const Type *Args[] = {
552 Type::getInt32PtrTy(*Ctx), // uint32_t *predecessor
553 Type::getInt64PtrTy(*Ctx)->getPointerTo(), // uint64_t **state_table_row
555 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
557 return M->getOrInsertFunction("llvm_gcda_increment_indirect_counter", FTy);
560 Constant *GCOVProfiler::getEmitFunctionFunc() {
561 const Type *Args[] = { Type::getInt32Ty(*Ctx) };
562 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
564 return M->getOrInsertFunction("llvm_gcda_emit_function", FTy);
567 Constant *GCOVProfiler::getEmitArcsFunc() {
568 const Type *Args[] = {
569 Type::getInt32Ty(*Ctx), // uint32_t num_counters
570 Type::getInt64PtrTy(*Ctx), // uint64_t *counters
572 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx),
574 return M->getOrInsertFunction("llvm_gcda_emit_arcs", FTy);
577 Constant *GCOVProfiler::getEndFileFunc() {
578 const FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false);
579 return M->getOrInsertFunction("llvm_gcda_end_file", FTy);
582 GlobalVariable *GCOVProfiler::getEdgeStateValue() {
583 GlobalVariable *GV = M->getGlobalVariable("__llvm_gcov_global_state_pred");
585 GV = new GlobalVariable(*M, Type::getInt32Ty(*Ctx), false,
586 GlobalValue::InternalLinkage,
587 ConstantInt::get(Type::getInt32Ty(*Ctx),
589 "__llvm_gcov_global_state_pred");
590 GV->setUnnamedAddr(true);
595 void GCOVProfiler::insertCounterWriteout(
596 DebugInfoFinder &DIF,
597 SmallVector<std::pair<GlobalVariable *, uint32_t>, 8> &CountersByIdent) {
598 const FunctionType *WriteoutFTy =
599 FunctionType::get(Type::getVoidTy(*Ctx), false);
600 Function *WriteoutF = Function::Create(WriteoutFTy,
601 GlobalValue::InternalLinkage,
602 "__llvm_gcov_writeout", M);
603 WriteoutF->setUnnamedAddr(true);
604 BasicBlock *BB = BasicBlock::Create(*Ctx, "", WriteoutF);
605 IRBuilder<> Builder(BB);
607 Constant *StartFile = getStartFileFunc();
608 Constant *EmitFunction = getEmitFunctionFunc();
609 Constant *EmitArcs = getEmitArcsFunc();
610 Constant *EndFile = getEndFileFunc();
612 for (DebugInfoFinder::iterator CUI = DIF.compile_unit_begin(),
613 CUE = DIF.compile_unit_end(); CUI != CUE; ++CUI) {
614 DICompileUnit compile_unit(*CUI);
615 std::string FilenameGcda = replaceStem(compile_unit.getFilename(), "gcda");
616 Builder.CreateCall(StartFile,
617 Builder.CreateGlobalStringPtr(FilenameGcda));
618 for (SmallVector<std::pair<GlobalVariable *, uint32_t>, 8>::iterator
619 I = CountersByIdent.begin(), E = CountersByIdent.end();
621 Builder.CreateCall(EmitFunction, ConstantInt::get(Type::getInt32Ty(*Ctx),
623 GlobalVariable *GV = I->first;
625 cast<ArrayType>(GV->getType()->getElementType())->getNumElements();
626 Builder.CreateCall2(EmitArcs,
627 ConstantInt::get(Type::getInt32Ty(*Ctx), Arcs),
628 Builder.CreateConstGEP2_64(GV, 0, 0));
630 Builder.CreateCall(EndFile);
632 Builder.CreateRetVoid();
634 InsertProfilingShutdownCall(WriteoutF, M);