1 //===- AddDiscriminators.cpp - Insert DWARF path discriminators -----------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file adds DWARF discriminators to the IR. Path discriminators are 11 // used to decide what CFG path was taken inside sub-graphs whose instructions 12 // share the same line and column number information. 13 // 14 // The main user of this is the sample profiler. Instruction samples are 15 // mapped to line number information. Since a single line may be spread 16 // out over several basic blocks, discriminators add more precise location 17 // for the samples. 18 // 19 // For example, 20 // 21 // 1 #define ASSERT(P) 22 // 2 if (!(P)) 23 // 3 abort() 24 // ... 25 // 100 while (true) { 26 // 101 ASSERT (sum < 0); 27 // 102 ... 28 // 130 } 29 // 30 // when converted to IR, this snippet looks something like: 31 // 32 // while.body: ; preds = %entry, %if.end 33 // %0 = load i32* %sum, align 4, !dbg !15 34 // %cmp = icmp slt i32 %0, 0, !dbg !15 35 // br i1 %cmp, label %if.end, label %if.then, !dbg !15 36 // 37 // if.then: ; preds = %while.body 38 // call void @abort(), !dbg !15 39 // br label %if.end, !dbg !15 40 // 41 // Notice that all the instructions in blocks 'while.body' and 'if.then' 42 // have exactly the same debug information. When this program is sampled 43 // at runtime, the profiler will assume that all these instructions are 44 // equally frequent. This, in turn, will consider the edge while.body->if.then 45 // to be frequently taken (which is incorrect). 46 // 47 // By adding a discriminator value to the instructions in block 'if.then', 48 // we can distinguish instructions at line 101 with discriminator 0 from 49 // the instructions at line 101 with discriminator 1. 50 // 51 // For more details about DWARF discriminators, please visit 52 // http://wiki.dwarfstd.org/index.php?title=Path_Discriminators 53 //===----------------------------------------------------------------------===// 54 55 #include "llvm/ADT/DenseMap.h" 56 #include "llvm/IR/BasicBlock.h" 57 #include "llvm/IR/Constants.h" 58 #include "llvm/IR/DIBuilder.h" 59 #include "llvm/IR/DebugInfo.h" 60 #include "llvm/IR/Instructions.h" 61 #include "llvm/IR/IntrinsicInst.h" 62 #include "llvm/IR/LLVMContext.h" 63 #include "llvm/IR/Module.h" 64 #include "llvm/Pass.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Debug.h" 67 #include "llvm/Support/raw_ostream.h" 68 #include "llvm/Transforms/Scalar.h" 69 70 using namespace llvm; 71 72 #define DEBUG_TYPE "add-discriminators" 73 74 namespace { 75 struct AddDiscriminators : public FunctionPass { 76 static char ID; // Pass identification, replacement for typeid 77 AddDiscriminators() : FunctionPass(ID) { 78 initializeAddDiscriminatorsPass(*PassRegistry::getPassRegistry()); 79 } 80 81 bool runOnFunction(Function &F) override; 82 }; 83 } 84 85 char AddDiscriminators::ID = 0; 86 INITIALIZE_PASS_BEGIN(AddDiscriminators, "add-discriminators", 87 "Add DWARF path discriminators", false, false) 88 INITIALIZE_PASS_END(AddDiscriminators, "add-discriminators", 89 "Add DWARF path discriminators", false, false) 90 91 // Command line option to disable discriminator generation even in the 92 // presence of debug information. This is only needed when debugging 93 // debug info generation issues. 94 static cl::opt<bool> NoDiscriminators( 95 "no-discriminators", cl::init(false), 96 cl::desc("Disable generation of discriminator information.")); 97 98 FunctionPass *llvm::createAddDiscriminatorsPass() { 99 return new AddDiscriminators(); 100 } 101 102 static bool hasDebugInfo(const Function &F) { 103 DISubprogram *S = getDISubprogram(&F); 104 return S != nullptr; 105 } 106 107 /// \brief Assign DWARF discriminators. 108 /// 109 /// To assign discriminators, we examine the boundaries of every 110 /// basic block and its successors. Suppose there is a basic block B1 111 /// with successor B2. The last instruction I1 in B1 and the first 112 /// instruction I2 in B2 are located at the same file and line number. 113 /// This situation is illustrated in the following code snippet: 114 /// 115 /// if (i < 10) x = i; 116 /// 117 /// entry: 118 /// br i1 %cmp, label %if.then, label %if.end, !dbg !10 119 /// if.then: 120 /// %1 = load i32* %i.addr, align 4, !dbg !10 121 /// store i32 %1, i32* %x, align 4, !dbg !10 122 /// br label %if.end, !dbg !10 123 /// if.end: 124 /// ret void, !dbg !12 125 /// 126 /// Notice how the branch instruction in block 'entry' and all the 127 /// instructions in block 'if.then' have the exact same debug location 128 /// information (!dbg !10). 129 /// 130 /// To distinguish instructions in block 'entry' from instructions in 131 /// block 'if.then', we generate a new lexical block for all the 132 /// instruction in block 'if.then' that share the same file and line 133 /// location with the last instruction of block 'entry'. 134 /// 135 /// This new lexical block will have the same location information as 136 /// the previous one, but with a new DWARF discriminator value. 137 /// 138 /// One of the main uses of this discriminator value is in runtime 139 /// sample profilers. It allows the profiler to distinguish instructions 140 /// at location !dbg !10 that execute on different basic blocks. This is 141 /// important because while the predicate 'if (x < 10)' may have been 142 /// executed millions of times, the assignment 'x = i' may have only 143 /// executed a handful of times (meaning that the entry->if.then edge is 144 /// seldom taken). 145 /// 146 /// If we did not have discriminator information, the profiler would 147 /// assign the same weight to both blocks 'entry' and 'if.then', which 148 /// in turn will make it conclude that the entry->if.then edge is very 149 /// hot. 150 /// 151 /// To decide where to create new discriminator values, this function 152 /// traverses the CFG and examines instruction at basic block boundaries. 153 /// If the last instruction I1 of a block B1 is at the same file and line 154 /// location as instruction I2 of successor B2, then it creates a new 155 /// lexical block for I2 and all the instruction in B2 that share the same 156 /// file and line location as I2. This new lexical block will have a 157 /// different discriminator number than I1. 158 bool AddDiscriminators::runOnFunction(Function &F) { 159 // If the function has debug information, but the user has disabled 160 // discriminators, do nothing. 161 // Simlarly, if the function has no debug info, do nothing. 162 // Finally, if this module is built with dwarf versions earlier than 4, 163 // do nothing (discriminator support is a DWARF 4 feature). 164 if (NoDiscriminators || !hasDebugInfo(F) || 165 F.getParent()->getDwarfVersion() < 4) 166 return false; 167 168 bool Changed = false; 169 Module *M = F.getParent(); 170 LLVMContext &Ctx = M->getContext(); 171 DIBuilder Builder(*M, /*AllowUnresolved*/ false); 172 173 typedef std::pair<StringRef, unsigned> Location; 174 typedef DenseMap<const BasicBlock *, Metadata *> BBScopeMap; 175 typedef DenseMap<Location, BBScopeMap> LocationBBMap; 176 177 LocationBBMap LBM; 178 179 // Traverse all instructions in the function. If the source line location 180 // of the instruction appears in other basic block, assign a new 181 // discriminator for this instruction. 182 for (BasicBlock &B : F) { 183 for (auto &I : B.getInstList()) { 184 if (isa<DbgInfoIntrinsic>(&I)) 185 continue; 186 const DILocation *DIL = I.getDebugLoc(); 187 if (!DIL) 188 continue; 189 Location L = std::make_pair(DIL->getFilename(), DIL->getLine()); 190 auto &BBMap = LBM[L]; 191 auto R = BBMap.insert(std::make_pair(&B, (Metadata *)nullptr)); 192 if (BBMap.size() == 1) 193 continue; 194 bool InsertSuccess = R.second; 195 Metadata *&NewScope = R.first->second; 196 // If we could insert a different block in the same location, a 197 // discriminator is needed to distinguish both instructions. 198 if (InsertSuccess) { 199 auto *Scope = DIL->getScope(); 200 auto *File = 201 Builder.createFile(DIL->getFilename(), Scope->getDirectory()); 202 NewScope = Builder.createLexicalBlockFile( 203 Scope, File, DIL->computeNewDiscriminator()); 204 } 205 I.setDebugLoc(DILocation::get(Ctx, DIL->getLine(), DIL->getColumn(), 206 NewScope, DIL->getInlinedAt())); 207 DEBUG(dbgs() << DIL->getFilename() << ":" << DIL->getLine() << ":" 208 << DIL->getColumn() << ":" 209 << dyn_cast<DILexicalBlockFile>(NewScope)->getDiscriminator() 210 << I << "\n"); 211 Changed = true; 212 } 213 } 214 215 // Traverse all instructions and assign new discriminators to call 216 // instructions with the same lineno that are in the same basic block. 217 // Sample base profile needs to distinguish different function calls within 218 // a same source line for correct profile annotation. 219 for (BasicBlock &B : F) { 220 const DILocation *FirstDIL = NULL; 221 for (auto &I : B.getInstList()) { 222 CallInst *Current = dyn_cast<CallInst>(&I); 223 if (!Current || isa<DbgInfoIntrinsic>(&I)) 224 continue; 225 226 DILocation *CurrentDIL = Current->getDebugLoc(); 227 if (FirstDIL) { 228 if (CurrentDIL && CurrentDIL->getLine() == FirstDIL->getLine() && 229 CurrentDIL->getFilename() == FirstDIL->getFilename()) { 230 auto *Scope = FirstDIL->getScope(); 231 auto *File = Builder.createFile(FirstDIL->getFilename(), 232 Scope->getDirectory()); 233 auto *NewScope = Builder.createLexicalBlockFile( 234 Scope, File, FirstDIL->computeNewDiscriminator()); 235 Current->setDebugLoc(DILocation::get( 236 Ctx, CurrentDIL->getLine(), CurrentDIL->getColumn(), NewScope, 237 CurrentDIL->getInlinedAt())); 238 Changed = true; 239 } else { 240 FirstDIL = CurrentDIL; 241 } 242 } else { 243 FirstDIL = CurrentDIL; 244 } 245 } 246 } 247 return Changed; 248 } 249