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/Scalar.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/LLVMContext.h" 62 #include "llvm/IR/Module.h" 63 #include "llvm/Pass.h" 64 #include "llvm/Support/CommandLine.h" 65 #include "llvm/Support/Debug.h" 66 #include "llvm/Support/raw_ostream.h" 67 68 using namespace llvm; 69 70 #define DEBUG_TYPE "add-discriminators" 71 72 namespace { 73 struct AddDiscriminators : public FunctionPass { 74 static char ID; // Pass identification, replacement for typeid 75 AddDiscriminators() : FunctionPass(ID) { 76 initializeAddDiscriminatorsPass(*PassRegistry::getPassRegistry()); 77 } 78 79 bool runOnFunction(Function &F) override; 80 }; 81 } 82 83 char AddDiscriminators::ID = 0; 84 INITIALIZE_PASS_BEGIN(AddDiscriminators, "add-discriminators", 85 "Add DWARF path discriminators", false, false) 86 INITIALIZE_PASS_END(AddDiscriminators, "add-discriminators", 87 "Add DWARF path discriminators", false, false) 88 89 // Command line option to disable discriminator generation even in the 90 // presence of debug information. This is only needed when debugging 91 // debug info generation issues. 92 static cl::opt<bool> 93 NoDiscriminators("no-discriminators", cl::init(false), 94 cl::desc("Disable generation of discriminator information.")); 95 96 FunctionPass *llvm::createAddDiscriminatorsPass() { 97 return new AddDiscriminators(); 98 } 99 100 static bool hasDebugInfo(const Function &F) { 101 NamedMDNode *CUNodes = F.getParent()->getNamedMetadata("llvm.dbg.cu"); 102 return CUNodes != nullptr; 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 bool AddDiscriminators::runOnFunction(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 // Finally, if this module is built with dwarf versions earlier than 4, 161 // do nothing (discriminator support is a DWARF 4 feature). 162 if (NoDiscriminators || 163 !hasDebugInfo(F) || 164 F.getParent()->getDwarfVersion() < 4) 165 return false; 166 167 bool Changed = false; 168 Module *M = F.getParent(); 169 LLVMContext &Ctx = M->getContext(); 170 DIBuilder Builder(*M, /*AllowUnresolved*/ false); 171 172 // Traverse all the blocks looking for instructions in different 173 // blocks that are at the same file:line location. 174 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) { 175 BasicBlock *B = I; 176 TerminatorInst *Last = B->getTerminator(); 177 DILocation LastDIL = Last->getDebugLoc().get(); 178 if (!LastDIL) 179 continue; 180 181 for (unsigned I = 0; I < Last->getNumSuccessors(); ++I) { 182 BasicBlock *Succ = Last->getSuccessor(I); 183 Instruction *First = Succ->getFirstNonPHIOrDbgOrLifetime(); 184 DILocation FirstDIL = First->getDebugLoc().get(); 185 if (!FirstDIL) 186 continue; 187 188 // If the first instruction (First) of Succ is at the same file 189 // location as B's last instruction (Last), add a new 190 // discriminator for First's location and all the instructions 191 // in Succ that share the same location with First. 192 if (!FirstDIL->canDiscriminate(*LastDIL)) { 193 // Create a new lexical scope and compute a new discriminator 194 // number for it. 195 StringRef Filename = FirstDIL->getFilename(); 196 auto *Scope = FirstDIL->getScope(); 197 DIFile File = Builder.createFile(Filename, Scope->getDirectory()); 198 199 // FIXME: Calculate the discriminator here, based on local information, 200 // and delete MDLocation::computeNewDiscriminator(). The current 201 // solution gives different results depending on other modules in the 202 // same context. All we really need is to discriminate between 203 // FirstDIL and LastDIL -- a local map would suffice. 204 unsigned Discriminator = FirstDIL->computeNewDiscriminator(); 205 DILexicalBlockFile NewScope = 206 Builder.createLexicalBlockFile(Scope, File, Discriminator); 207 auto *NewDIL = 208 MDLocation::get(Ctx, FirstDIL->getLine(), FirstDIL->getColumn(), 209 NewScope, FirstDIL->getInlinedAt()); 210 DebugLoc newDebugLoc = NewDIL; 211 212 // Attach this new debug location to First and every 213 // instruction following First that shares the same location. 214 for (BasicBlock::iterator I1(*First), E1 = Succ->end(); I1 != E1; 215 ++I1) { 216 if (I1->getDebugLoc().get() != FirstDIL) 217 break; 218 I1->setDebugLoc(newDebugLoc); 219 DEBUG(dbgs() << NewDIL->getFilename() << ":" << NewDIL->getLine() 220 << ":" << NewDIL->getColumn() << ":" 221 << NewDIL->getDiscriminator() << *I1 << "\n"); 222 } 223 DEBUG(dbgs() << "\n"); 224 Changed = true; 225 } 226 } 227 } 228 return Changed; 229 } 230