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/Transforms/Utils/AddDiscriminators.h" 56 #include "llvm/ADT/DenseMap.h" 57 #include "llvm/ADT/DenseSet.h" 58 #include "llvm/IR/BasicBlock.h" 59 #include "llvm/IR/Constants.h" 60 #include "llvm/IR/DebugInfo.h" 61 #include "llvm/IR/Instructions.h" 62 #include "llvm/IR/IntrinsicInst.h" 63 #include "llvm/IR/LLVMContext.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 // The legacy pass of AddDiscriminators. 76 struct AddDiscriminatorsLegacyPass : public FunctionPass { 77 static char ID; // Pass identification, replacement for typeid 78 AddDiscriminatorsLegacyPass() : FunctionPass(ID) { 79 initializeAddDiscriminatorsLegacyPassPass(*PassRegistry::getPassRegistry()); 80 } 81 82 bool runOnFunction(Function &F) override; 83 }; 84 85 } // end anonymous namespace 86 87 char AddDiscriminatorsLegacyPass::ID = 0; 88 INITIALIZE_PASS_BEGIN(AddDiscriminatorsLegacyPass, "add-discriminators", 89 "Add DWARF path discriminators", false, false) 90 INITIALIZE_PASS_END(AddDiscriminatorsLegacyPass, "add-discriminators", 91 "Add DWARF path discriminators", false, false) 92 93 // Command line option to disable discriminator generation even in the 94 // presence of debug information. This is only needed when debugging 95 // debug info generation issues. 96 static cl::opt<bool> NoDiscriminators( 97 "no-discriminators", cl::init(false), 98 cl::desc("Disable generation of discriminator information.")); 99 100 // Create the legacy AddDiscriminatorsPass. 101 FunctionPass *llvm::createAddDiscriminatorsPass() { 102 return new AddDiscriminatorsLegacyPass(); 103 } 104 105 /// \brief Assign DWARF discriminators. 106 /// 107 /// To assign discriminators, we examine the boundaries of every 108 /// basic block and its successors. Suppose there is a basic block B1 109 /// with successor B2. The last instruction I1 in B1 and the first 110 /// instruction I2 in B2 are located at the same file and line number. 111 /// This situation is illustrated in the following code snippet: 112 /// 113 /// if (i < 10) x = i; 114 /// 115 /// entry: 116 /// br i1 %cmp, label %if.then, label %if.end, !dbg !10 117 /// if.then: 118 /// %1 = load i32* %i.addr, align 4, !dbg !10 119 /// store i32 %1, i32* %x, align 4, !dbg !10 120 /// br label %if.end, !dbg !10 121 /// if.end: 122 /// ret void, !dbg !12 123 /// 124 /// Notice how the branch instruction in block 'entry' and all the 125 /// instructions in block 'if.then' have the exact same debug location 126 /// information (!dbg !10). 127 /// 128 /// To distinguish instructions in block 'entry' from instructions in 129 /// block 'if.then', we generate a new lexical block for all the 130 /// instruction in block 'if.then' that share the same file and line 131 /// location with the last instruction of block 'entry'. 132 /// 133 /// This new lexical block will have the same location information as 134 /// the previous one, but with a new DWARF discriminator value. 135 /// 136 /// One of the main uses of this discriminator value is in runtime 137 /// sample profilers. It allows the profiler to distinguish instructions 138 /// at location !dbg !10 that execute on different basic blocks. This is 139 /// important because while the predicate 'if (x < 10)' may have been 140 /// executed millions of times, the assignment 'x = i' may have only 141 /// executed a handful of times (meaning that the entry->if.then edge is 142 /// seldom taken). 143 /// 144 /// If we did not have discriminator information, the profiler would 145 /// assign the same weight to both blocks 'entry' and 'if.then', which 146 /// in turn will make it conclude that the entry->if.then edge is very 147 /// hot. 148 /// 149 /// To decide where to create new discriminator values, this function 150 /// traverses the CFG and examines instruction at basic block boundaries. 151 /// If the last instruction I1 of a block B1 is at the same file and line 152 /// location as instruction I2 of successor B2, then it creates a new 153 /// lexical block for I2 and all the instruction in B2 that share the same 154 /// file and line location as I2. This new lexical block will have a 155 /// different discriminator number than I1. 156 static bool addDiscriminators(Function &F) { 157 // If the function has debug information, but the user has disabled 158 // discriminators, do nothing. 159 // Simlarly, if the function has no debug info, do nothing. 160 if (NoDiscriminators || !F.getSubprogram()) 161 return false; 162 163 bool Changed = false; 164 165 typedef std::pair<StringRef, unsigned> Location; 166 typedef DenseSet<const BasicBlock *> BBSet; 167 typedef DenseMap<Location, BBSet> LocationBBMap; 168 typedef DenseMap<Location, unsigned> LocationDiscriminatorMap; 169 typedef DenseSet<Location> LocationSet; 170 171 LocationBBMap LBM; 172 LocationDiscriminatorMap LDM; 173 174 // Traverse all instructions in the function. If the source line location 175 // of the instruction appears in other basic block, assign a new 176 // discriminator for this instruction. 177 for (BasicBlock &B : F) { 178 for (auto &I : B.getInstList()) { 179 if (isa<IntrinsicInst>(&I)) 180 continue; 181 const DILocation *DIL = I.getDebugLoc(); 182 if (!DIL) 183 continue; 184 Location L = std::make_pair(DIL->getFilename(), DIL->getLine()); 185 auto &BBMap = LBM[L]; 186 auto R = BBMap.insert(&B); 187 if (BBMap.size() == 1) 188 continue; 189 // If we could insert more than one block with the same line+file, a 190 // discriminator is needed to distinguish both instructions. 191 unsigned Discriminator = R.second ? ++LDM[L] : LDM[L]; 192 I.setDebugLoc(DIL->cloneWithDiscriminator(Discriminator)); 193 DEBUG(dbgs() << DIL->getFilename() << ":" << DIL->getLine() << ":" 194 << DIL->getColumn() << ":" 195 << Discriminator << " " 196 << I << "\n"); 197 Changed = true; 198 } 199 } 200 201 // Traverse all instructions and assign new discriminators to call 202 // instructions with the same lineno that are in the same basic block. 203 // Sample base profile needs to distinguish different function calls within 204 // a same source line for correct profile annotation. 205 for (BasicBlock &B : F) { 206 LocationSet CallLocations; 207 for (auto &I : B.getInstList()) { 208 CallInst *Current = dyn_cast<CallInst>(&I); 209 if (!Current || isa<IntrinsicInst>(&I)) 210 continue; 211 212 DILocation *CurrentDIL = Current->getDebugLoc(); 213 if (!CurrentDIL) 214 continue; 215 Location L = 216 std::make_pair(CurrentDIL->getFilename(), CurrentDIL->getLine()); 217 if (!CallLocations.insert(L).second) { 218 Current->setDebugLoc(CurrentDIL->cloneWithDiscriminator(++LDM[L])); 219 Changed = true; 220 } 221 } 222 } 223 return Changed; 224 } 225 226 bool AddDiscriminatorsLegacyPass::runOnFunction(Function &F) { 227 return addDiscriminators(F); 228 } 229 PreservedAnalyses AddDiscriminatorsPass::run(Function &F, 230 FunctionAnalysisManager &AM) { 231 if (!addDiscriminators(F)) 232 return PreservedAnalyses::all(); 233 234 // FIXME: should be all() 235 return PreservedAnalyses::none(); 236 } 237