1e8d8bef9SDimitry Andric //===- MemProfiler.cpp - memory allocation and access profiler ------------===// 2e8d8bef9SDimitry Andric // 3e8d8bef9SDimitry Andric // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4e8d8bef9SDimitry Andric // See https://llvm.org/LICENSE.txt for license information. 5e8d8bef9SDimitry Andric // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6e8d8bef9SDimitry Andric // 7e8d8bef9SDimitry Andric //===----------------------------------------------------------------------===// 8e8d8bef9SDimitry Andric // 9e8d8bef9SDimitry Andric // This file is a part of MemProfiler. Memory accesses are instrumented 10e8d8bef9SDimitry Andric // to increment the access count held in a shadow memory location, or 11e8d8bef9SDimitry Andric // alternatively to call into the runtime. Memory intrinsic calls (memmove, 12e8d8bef9SDimitry Andric // memcpy, memset) are changed to call the memory profiling runtime version 13e8d8bef9SDimitry Andric // instead. 14e8d8bef9SDimitry Andric // 15e8d8bef9SDimitry Andric //===----------------------------------------------------------------------===// 16e8d8bef9SDimitry Andric 17e8d8bef9SDimitry Andric #include "llvm/Transforms/Instrumentation/MemProfiler.h" 18e8d8bef9SDimitry Andric #include "llvm/ADT/SmallVector.h" 19e8d8bef9SDimitry Andric #include "llvm/ADT/Statistic.h" 20e8d8bef9SDimitry Andric #include "llvm/ADT/StringRef.h" 21e8d8bef9SDimitry Andric #include "llvm/ADT/Triple.h" 22349cc55cSDimitry Andric #include "llvm/Analysis/ValueTracking.h" 23e8d8bef9SDimitry Andric #include "llvm/IR/Constant.h" 24e8d8bef9SDimitry Andric #include "llvm/IR/DataLayout.h" 25e8d8bef9SDimitry Andric #include "llvm/IR/Function.h" 26e8d8bef9SDimitry Andric #include "llvm/IR/GlobalValue.h" 27e8d8bef9SDimitry Andric #include "llvm/IR/IRBuilder.h" 28e8d8bef9SDimitry Andric #include "llvm/IR/Instruction.h" 291fd87a68SDimitry Andric #include "llvm/IR/IntrinsicInst.h" 30e8d8bef9SDimitry Andric #include "llvm/IR/Module.h" 31e8d8bef9SDimitry Andric #include "llvm/IR/Type.h" 32e8d8bef9SDimitry Andric #include "llvm/IR/Value.h" 33e8d8bef9SDimitry Andric #include "llvm/InitializePasses.h" 34e8d8bef9SDimitry Andric #include "llvm/Pass.h" 3581ad6265SDimitry Andric #include "llvm/ProfileData/InstrProf.h" 36e8d8bef9SDimitry Andric #include "llvm/Support/CommandLine.h" 37e8d8bef9SDimitry Andric #include "llvm/Support/Debug.h" 38e8d8bef9SDimitry Andric #include "llvm/Transforms/Utils/BasicBlockUtils.h" 39e8d8bef9SDimitry Andric #include "llvm/Transforms/Utils/ModuleUtils.h" 40e8d8bef9SDimitry Andric 41e8d8bef9SDimitry Andric using namespace llvm; 42e8d8bef9SDimitry Andric 43e8d8bef9SDimitry Andric #define DEBUG_TYPE "memprof" 44e8d8bef9SDimitry Andric 45e8d8bef9SDimitry Andric constexpr int LLVM_MEM_PROFILER_VERSION = 1; 46e8d8bef9SDimitry Andric 47e8d8bef9SDimitry Andric // Size of memory mapped to a single shadow location. 48e8d8bef9SDimitry Andric constexpr uint64_t DefaultShadowGranularity = 64; 49e8d8bef9SDimitry Andric 50e8d8bef9SDimitry Andric // Scale from granularity down to shadow size. 51e8d8bef9SDimitry Andric constexpr uint64_t DefaultShadowScale = 3; 52e8d8bef9SDimitry Andric 53e8d8bef9SDimitry Andric constexpr char MemProfModuleCtorName[] = "memprof.module_ctor"; 54e8d8bef9SDimitry Andric constexpr uint64_t MemProfCtorAndDtorPriority = 1; 55e8d8bef9SDimitry Andric // On Emscripten, the system needs more than one priorities for constructors. 56e8d8bef9SDimitry Andric constexpr uint64_t MemProfEmscriptenCtorAndDtorPriority = 50; 57e8d8bef9SDimitry Andric constexpr char MemProfInitName[] = "__memprof_init"; 58e8d8bef9SDimitry Andric constexpr char MemProfVersionCheckNamePrefix[] = 59e8d8bef9SDimitry Andric "__memprof_version_mismatch_check_v"; 60e8d8bef9SDimitry Andric 61e8d8bef9SDimitry Andric constexpr char MemProfShadowMemoryDynamicAddress[] = 62e8d8bef9SDimitry Andric "__memprof_shadow_memory_dynamic_address"; 63e8d8bef9SDimitry Andric 64e8d8bef9SDimitry Andric constexpr char MemProfFilenameVar[] = "__memprof_profile_filename"; 65e8d8bef9SDimitry Andric 66e8d8bef9SDimitry Andric // Command-line flags. 67e8d8bef9SDimitry Andric 68e8d8bef9SDimitry Andric static cl::opt<bool> ClInsertVersionCheck( 69e8d8bef9SDimitry Andric "memprof-guard-against-version-mismatch", 70e8d8bef9SDimitry Andric cl::desc("Guard against compiler/runtime version mismatch."), cl::Hidden, 71e8d8bef9SDimitry Andric cl::init(true)); 72e8d8bef9SDimitry Andric 73e8d8bef9SDimitry Andric // This flag may need to be replaced with -f[no-]memprof-reads. 74e8d8bef9SDimitry Andric static cl::opt<bool> ClInstrumentReads("memprof-instrument-reads", 75e8d8bef9SDimitry Andric cl::desc("instrument read instructions"), 76e8d8bef9SDimitry Andric cl::Hidden, cl::init(true)); 77e8d8bef9SDimitry Andric 78e8d8bef9SDimitry Andric static cl::opt<bool> 79e8d8bef9SDimitry Andric ClInstrumentWrites("memprof-instrument-writes", 80e8d8bef9SDimitry Andric cl::desc("instrument write instructions"), cl::Hidden, 81e8d8bef9SDimitry Andric cl::init(true)); 82e8d8bef9SDimitry Andric 83e8d8bef9SDimitry Andric static cl::opt<bool> ClInstrumentAtomics( 84e8d8bef9SDimitry Andric "memprof-instrument-atomics", 85e8d8bef9SDimitry Andric cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden, 86e8d8bef9SDimitry Andric cl::init(true)); 87e8d8bef9SDimitry Andric 88e8d8bef9SDimitry Andric static cl::opt<bool> ClUseCalls( 89e8d8bef9SDimitry Andric "memprof-use-callbacks", 90e8d8bef9SDimitry Andric cl::desc("Use callbacks instead of inline instrumentation sequences."), 91e8d8bef9SDimitry Andric cl::Hidden, cl::init(false)); 92e8d8bef9SDimitry Andric 93e8d8bef9SDimitry Andric static cl::opt<std::string> 94e8d8bef9SDimitry Andric ClMemoryAccessCallbackPrefix("memprof-memory-access-callback-prefix", 95e8d8bef9SDimitry Andric cl::desc("Prefix for memory access callbacks"), 96e8d8bef9SDimitry Andric cl::Hidden, cl::init("__memprof_")); 97e8d8bef9SDimitry Andric 98e8d8bef9SDimitry Andric // These flags allow to change the shadow mapping. 99e8d8bef9SDimitry Andric // The shadow mapping looks like 100e8d8bef9SDimitry Andric // Shadow = ((Mem & mask) >> scale) + offset 101e8d8bef9SDimitry Andric 102e8d8bef9SDimitry Andric static cl::opt<int> ClMappingScale("memprof-mapping-scale", 103e8d8bef9SDimitry Andric cl::desc("scale of memprof shadow mapping"), 104e8d8bef9SDimitry Andric cl::Hidden, cl::init(DefaultShadowScale)); 105e8d8bef9SDimitry Andric 106e8d8bef9SDimitry Andric static cl::opt<int> 107e8d8bef9SDimitry Andric ClMappingGranularity("memprof-mapping-granularity", 108e8d8bef9SDimitry Andric cl::desc("granularity of memprof shadow mapping"), 109e8d8bef9SDimitry Andric cl::Hidden, cl::init(DefaultShadowGranularity)); 110e8d8bef9SDimitry Andric 111349cc55cSDimitry Andric static cl::opt<bool> ClStack("memprof-instrument-stack", 112349cc55cSDimitry Andric cl::desc("Instrument scalar stack variables"), 113349cc55cSDimitry Andric cl::Hidden, cl::init(false)); 114349cc55cSDimitry Andric 115e8d8bef9SDimitry Andric // Debug flags. 116e8d8bef9SDimitry Andric 117e8d8bef9SDimitry Andric static cl::opt<int> ClDebug("memprof-debug", cl::desc("debug"), cl::Hidden, 118e8d8bef9SDimitry Andric cl::init(0)); 119e8d8bef9SDimitry Andric 120e8d8bef9SDimitry Andric static cl::opt<std::string> ClDebugFunc("memprof-debug-func", cl::Hidden, 121e8d8bef9SDimitry Andric cl::desc("Debug func")); 122e8d8bef9SDimitry Andric 123e8d8bef9SDimitry Andric static cl::opt<int> ClDebugMin("memprof-debug-min", cl::desc("Debug min inst"), 124e8d8bef9SDimitry Andric cl::Hidden, cl::init(-1)); 125e8d8bef9SDimitry Andric 126e8d8bef9SDimitry Andric static cl::opt<int> ClDebugMax("memprof-debug-max", cl::desc("Debug max inst"), 127e8d8bef9SDimitry Andric cl::Hidden, cl::init(-1)); 128e8d8bef9SDimitry Andric 129e8d8bef9SDimitry Andric STATISTIC(NumInstrumentedReads, "Number of instrumented reads"); 130e8d8bef9SDimitry Andric STATISTIC(NumInstrumentedWrites, "Number of instrumented writes"); 131349cc55cSDimitry Andric STATISTIC(NumSkippedStackReads, "Number of non-instrumented stack reads"); 132349cc55cSDimitry Andric STATISTIC(NumSkippedStackWrites, "Number of non-instrumented stack writes"); 133e8d8bef9SDimitry Andric 134e8d8bef9SDimitry Andric namespace { 135e8d8bef9SDimitry Andric 136e8d8bef9SDimitry Andric /// This struct defines the shadow mapping using the rule: 137e8d8bef9SDimitry Andric /// shadow = ((mem & mask) >> Scale) ADD DynamicShadowOffset. 138e8d8bef9SDimitry Andric struct ShadowMapping { 139e8d8bef9SDimitry Andric ShadowMapping() { 140e8d8bef9SDimitry Andric Scale = ClMappingScale; 141e8d8bef9SDimitry Andric Granularity = ClMappingGranularity; 142e8d8bef9SDimitry Andric Mask = ~(Granularity - 1); 143e8d8bef9SDimitry Andric } 144e8d8bef9SDimitry Andric 145e8d8bef9SDimitry Andric int Scale; 146e8d8bef9SDimitry Andric int Granularity; 147e8d8bef9SDimitry Andric uint64_t Mask; // Computed as ~(Granularity-1) 148e8d8bef9SDimitry Andric }; 149e8d8bef9SDimitry Andric 150e8d8bef9SDimitry Andric static uint64_t getCtorAndDtorPriority(Triple &TargetTriple) { 151e8d8bef9SDimitry Andric return TargetTriple.isOSEmscripten() ? MemProfEmscriptenCtorAndDtorPriority 152e8d8bef9SDimitry Andric : MemProfCtorAndDtorPriority; 153e8d8bef9SDimitry Andric } 154e8d8bef9SDimitry Andric 155e8d8bef9SDimitry Andric struct InterestingMemoryAccess { 156e8d8bef9SDimitry Andric Value *Addr = nullptr; 157e8d8bef9SDimitry Andric bool IsWrite; 15804eeddc0SDimitry Andric Type *AccessTy; 159e8d8bef9SDimitry Andric uint64_t TypeSize; 160e8d8bef9SDimitry Andric Value *MaybeMask = nullptr; 161e8d8bef9SDimitry Andric }; 162e8d8bef9SDimitry Andric 163e8d8bef9SDimitry Andric /// Instrument the code in module to profile memory accesses. 164e8d8bef9SDimitry Andric class MemProfiler { 165e8d8bef9SDimitry Andric public: 166e8d8bef9SDimitry Andric MemProfiler(Module &M) { 167e8d8bef9SDimitry Andric C = &(M.getContext()); 168e8d8bef9SDimitry Andric LongSize = M.getDataLayout().getPointerSizeInBits(); 169e8d8bef9SDimitry Andric IntptrTy = Type::getIntNTy(*C, LongSize); 170e8d8bef9SDimitry Andric } 171e8d8bef9SDimitry Andric 172e8d8bef9SDimitry Andric /// If it is an interesting memory access, populate information 173e8d8bef9SDimitry Andric /// about the access and return a InterestingMemoryAccess struct. 174*bdd1243dSDimitry Andric /// Otherwise return std::nullopt. 175*bdd1243dSDimitry Andric std::optional<InterestingMemoryAccess> 176e8d8bef9SDimitry Andric isInterestingMemoryAccess(Instruction *I) const; 177e8d8bef9SDimitry Andric 178e8d8bef9SDimitry Andric void instrumentMop(Instruction *I, const DataLayout &DL, 179e8d8bef9SDimitry Andric InterestingMemoryAccess &Access); 180e8d8bef9SDimitry Andric void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore, 181e8d8bef9SDimitry Andric Value *Addr, uint32_t TypeSize, bool IsWrite); 182e8d8bef9SDimitry Andric void instrumentMaskedLoadOrStore(const DataLayout &DL, Value *Mask, 18381ad6265SDimitry Andric Instruction *I, Value *Addr, Type *AccessTy, 184e8d8bef9SDimitry Andric bool IsWrite); 185e8d8bef9SDimitry Andric void instrumentMemIntrinsic(MemIntrinsic *MI); 186e8d8bef9SDimitry Andric Value *memToShadow(Value *Shadow, IRBuilder<> &IRB); 187e8d8bef9SDimitry Andric bool instrumentFunction(Function &F); 188e8d8bef9SDimitry Andric bool maybeInsertMemProfInitAtFunctionEntry(Function &F); 189e8d8bef9SDimitry Andric bool insertDynamicShadowAtFunctionEntry(Function &F); 190e8d8bef9SDimitry Andric 191e8d8bef9SDimitry Andric private: 192e8d8bef9SDimitry Andric void initializeCallbacks(Module &M); 193e8d8bef9SDimitry Andric 194e8d8bef9SDimitry Andric LLVMContext *C; 195e8d8bef9SDimitry Andric int LongSize; 196e8d8bef9SDimitry Andric Type *IntptrTy; 197e8d8bef9SDimitry Andric ShadowMapping Mapping; 198e8d8bef9SDimitry Andric 199e8d8bef9SDimitry Andric // These arrays is indexed by AccessIsWrite 200e8d8bef9SDimitry Andric FunctionCallee MemProfMemoryAccessCallback[2]; 201e8d8bef9SDimitry Andric FunctionCallee MemProfMemoryAccessCallbackSized[2]; 202e8d8bef9SDimitry Andric 203e8d8bef9SDimitry Andric FunctionCallee MemProfMemmove, MemProfMemcpy, MemProfMemset; 204e8d8bef9SDimitry Andric Value *DynamicShadowOffset = nullptr; 205e8d8bef9SDimitry Andric }; 206e8d8bef9SDimitry Andric 207e8d8bef9SDimitry Andric class ModuleMemProfiler { 208e8d8bef9SDimitry Andric public: 209e8d8bef9SDimitry Andric ModuleMemProfiler(Module &M) { TargetTriple = Triple(M.getTargetTriple()); } 210e8d8bef9SDimitry Andric 211e8d8bef9SDimitry Andric bool instrumentModule(Module &); 212e8d8bef9SDimitry Andric 213e8d8bef9SDimitry Andric private: 214e8d8bef9SDimitry Andric Triple TargetTriple; 215e8d8bef9SDimitry Andric ShadowMapping Mapping; 216e8d8bef9SDimitry Andric Function *MemProfCtorFunction = nullptr; 217e8d8bef9SDimitry Andric }; 218e8d8bef9SDimitry Andric 219e8d8bef9SDimitry Andric } // end anonymous namespace 220e8d8bef9SDimitry Andric 22181ad6265SDimitry Andric MemProfilerPass::MemProfilerPass() = default; 222e8d8bef9SDimitry Andric 223e8d8bef9SDimitry Andric PreservedAnalyses MemProfilerPass::run(Function &F, 224e8d8bef9SDimitry Andric AnalysisManager<Function> &AM) { 225e8d8bef9SDimitry Andric Module &M = *F.getParent(); 226e8d8bef9SDimitry Andric MemProfiler Profiler(M); 227e8d8bef9SDimitry Andric if (Profiler.instrumentFunction(F)) 228e8d8bef9SDimitry Andric return PreservedAnalyses::none(); 229e8d8bef9SDimitry Andric return PreservedAnalyses::all(); 230e8d8bef9SDimitry Andric } 231e8d8bef9SDimitry Andric 23281ad6265SDimitry Andric ModuleMemProfilerPass::ModuleMemProfilerPass() = default; 233e8d8bef9SDimitry Andric 234e8d8bef9SDimitry Andric PreservedAnalyses ModuleMemProfilerPass::run(Module &M, 235e8d8bef9SDimitry Andric AnalysisManager<Module> &AM) { 236e8d8bef9SDimitry Andric ModuleMemProfiler Profiler(M); 237e8d8bef9SDimitry Andric if (Profiler.instrumentModule(M)) 238e8d8bef9SDimitry Andric return PreservedAnalyses::none(); 239e8d8bef9SDimitry Andric return PreservedAnalyses::all(); 240e8d8bef9SDimitry Andric } 241e8d8bef9SDimitry Andric 242e8d8bef9SDimitry Andric Value *MemProfiler::memToShadow(Value *Shadow, IRBuilder<> &IRB) { 243e8d8bef9SDimitry Andric // (Shadow & mask) >> scale 244e8d8bef9SDimitry Andric Shadow = IRB.CreateAnd(Shadow, Mapping.Mask); 245e8d8bef9SDimitry Andric Shadow = IRB.CreateLShr(Shadow, Mapping.Scale); 246e8d8bef9SDimitry Andric // (Shadow >> scale) | offset 247e8d8bef9SDimitry Andric assert(DynamicShadowOffset); 248e8d8bef9SDimitry Andric return IRB.CreateAdd(Shadow, DynamicShadowOffset); 249e8d8bef9SDimitry Andric } 250e8d8bef9SDimitry Andric 251e8d8bef9SDimitry Andric // Instrument memset/memmove/memcpy 252e8d8bef9SDimitry Andric void MemProfiler::instrumentMemIntrinsic(MemIntrinsic *MI) { 253e8d8bef9SDimitry Andric IRBuilder<> IRB(MI); 254e8d8bef9SDimitry Andric if (isa<MemTransferInst>(MI)) { 255e8d8bef9SDimitry Andric IRB.CreateCall( 256e8d8bef9SDimitry Andric isa<MemMoveInst>(MI) ? MemProfMemmove : MemProfMemcpy, 257e8d8bef9SDimitry Andric {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 258e8d8bef9SDimitry Andric IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()), 259e8d8bef9SDimitry Andric IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 260e8d8bef9SDimitry Andric } else if (isa<MemSetInst>(MI)) { 261e8d8bef9SDimitry Andric IRB.CreateCall( 262e8d8bef9SDimitry Andric MemProfMemset, 263e8d8bef9SDimitry Andric {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 264e8d8bef9SDimitry Andric IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false), 265e8d8bef9SDimitry Andric IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 266e8d8bef9SDimitry Andric } 267e8d8bef9SDimitry Andric MI->eraseFromParent(); 268e8d8bef9SDimitry Andric } 269e8d8bef9SDimitry Andric 270*bdd1243dSDimitry Andric std::optional<InterestingMemoryAccess> 271e8d8bef9SDimitry Andric MemProfiler::isInterestingMemoryAccess(Instruction *I) const { 272e8d8bef9SDimitry Andric // Do not instrument the load fetching the dynamic shadow address. 273e8d8bef9SDimitry Andric if (DynamicShadowOffset == I) 274*bdd1243dSDimitry Andric return std::nullopt; 275e8d8bef9SDimitry Andric 276e8d8bef9SDimitry Andric InterestingMemoryAccess Access; 277e8d8bef9SDimitry Andric 278e8d8bef9SDimitry Andric if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 279e8d8bef9SDimitry Andric if (!ClInstrumentReads) 280*bdd1243dSDimitry Andric return std::nullopt; 281e8d8bef9SDimitry Andric Access.IsWrite = false; 28204eeddc0SDimitry Andric Access.AccessTy = LI->getType(); 283e8d8bef9SDimitry Andric Access.Addr = LI->getPointerOperand(); 284e8d8bef9SDimitry Andric } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 285e8d8bef9SDimitry Andric if (!ClInstrumentWrites) 286*bdd1243dSDimitry Andric return std::nullopt; 287e8d8bef9SDimitry Andric Access.IsWrite = true; 28804eeddc0SDimitry Andric Access.AccessTy = SI->getValueOperand()->getType(); 289e8d8bef9SDimitry Andric Access.Addr = SI->getPointerOperand(); 290e8d8bef9SDimitry Andric } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { 291e8d8bef9SDimitry Andric if (!ClInstrumentAtomics) 292*bdd1243dSDimitry Andric return std::nullopt; 293e8d8bef9SDimitry Andric Access.IsWrite = true; 29404eeddc0SDimitry Andric Access.AccessTy = RMW->getValOperand()->getType(); 295e8d8bef9SDimitry Andric Access.Addr = RMW->getPointerOperand(); 296e8d8bef9SDimitry Andric } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) { 297e8d8bef9SDimitry Andric if (!ClInstrumentAtomics) 298*bdd1243dSDimitry Andric return std::nullopt; 299e8d8bef9SDimitry Andric Access.IsWrite = true; 30004eeddc0SDimitry Andric Access.AccessTy = XCHG->getCompareOperand()->getType(); 301e8d8bef9SDimitry Andric Access.Addr = XCHG->getPointerOperand(); 302e8d8bef9SDimitry Andric } else if (auto *CI = dyn_cast<CallInst>(I)) { 303e8d8bef9SDimitry Andric auto *F = CI->getCalledFunction(); 304e8d8bef9SDimitry Andric if (F && (F->getIntrinsicID() == Intrinsic::masked_load || 305e8d8bef9SDimitry Andric F->getIntrinsicID() == Intrinsic::masked_store)) { 306e8d8bef9SDimitry Andric unsigned OpOffset = 0; 307e8d8bef9SDimitry Andric if (F->getIntrinsicID() == Intrinsic::masked_store) { 308e8d8bef9SDimitry Andric if (!ClInstrumentWrites) 309*bdd1243dSDimitry Andric return std::nullopt; 310e8d8bef9SDimitry Andric // Masked store has an initial operand for the value. 311e8d8bef9SDimitry Andric OpOffset = 1; 31204eeddc0SDimitry Andric Access.AccessTy = CI->getArgOperand(0)->getType(); 313e8d8bef9SDimitry Andric Access.IsWrite = true; 314e8d8bef9SDimitry Andric } else { 315e8d8bef9SDimitry Andric if (!ClInstrumentReads) 316*bdd1243dSDimitry Andric return std::nullopt; 31704eeddc0SDimitry Andric Access.AccessTy = CI->getType(); 318e8d8bef9SDimitry Andric Access.IsWrite = false; 319e8d8bef9SDimitry Andric } 320e8d8bef9SDimitry Andric 321e8d8bef9SDimitry Andric auto *BasePtr = CI->getOperand(0 + OpOffset); 322e8d8bef9SDimitry Andric Access.MaybeMask = CI->getOperand(2 + OpOffset); 323e8d8bef9SDimitry Andric Access.Addr = BasePtr; 324e8d8bef9SDimitry Andric } 325e8d8bef9SDimitry Andric } 326e8d8bef9SDimitry Andric 327e8d8bef9SDimitry Andric if (!Access.Addr) 328*bdd1243dSDimitry Andric return std::nullopt; 329e8d8bef9SDimitry Andric 330*bdd1243dSDimitry Andric // Do not instrument accesses from different address spaces; we cannot deal 331e8d8bef9SDimitry Andric // with them. 332e8d8bef9SDimitry Andric Type *PtrTy = cast<PointerType>(Access.Addr->getType()->getScalarType()); 333e8d8bef9SDimitry Andric if (PtrTy->getPointerAddressSpace() != 0) 334*bdd1243dSDimitry Andric return std::nullopt; 335e8d8bef9SDimitry Andric 336e8d8bef9SDimitry Andric // Ignore swifterror addresses. 337e8d8bef9SDimitry Andric // swifterror memory addresses are mem2reg promoted by instruction 338e8d8bef9SDimitry Andric // selection. As such they cannot have regular uses like an instrumentation 339e8d8bef9SDimitry Andric // function and it makes no sense to track them as memory. 340e8d8bef9SDimitry Andric if (Access.Addr->isSwiftError()) 341*bdd1243dSDimitry Andric return std::nullopt; 342e8d8bef9SDimitry Andric 34381ad6265SDimitry Andric // Peel off GEPs and BitCasts. 34481ad6265SDimitry Andric auto *Addr = Access.Addr->stripInBoundsOffsets(); 34581ad6265SDimitry Andric 34681ad6265SDimitry Andric if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 34781ad6265SDimitry Andric // Do not instrument PGO counter updates. 34881ad6265SDimitry Andric if (GV->hasSection()) { 34981ad6265SDimitry Andric StringRef SectionName = GV->getSection(); 35081ad6265SDimitry Andric // Check if the global is in the PGO counters section. 35181ad6265SDimitry Andric auto OF = Triple(I->getModule()->getTargetTriple()).getObjectFormat(); 35281ad6265SDimitry Andric if (SectionName.endswith( 35381ad6265SDimitry Andric getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false))) 354*bdd1243dSDimitry Andric return std::nullopt; 35581ad6265SDimitry Andric } 35681ad6265SDimitry Andric 35781ad6265SDimitry Andric // Do not instrument accesses to LLVM internal variables. 35881ad6265SDimitry Andric if (GV->getName().startswith("__llvm")) 359*bdd1243dSDimitry Andric return std::nullopt; 36081ad6265SDimitry Andric } 36181ad6265SDimitry Andric 36204eeddc0SDimitry Andric const DataLayout &DL = I->getModule()->getDataLayout(); 36304eeddc0SDimitry Andric Access.TypeSize = DL.getTypeStoreSizeInBits(Access.AccessTy); 364e8d8bef9SDimitry Andric return Access; 365e8d8bef9SDimitry Andric } 366e8d8bef9SDimitry Andric 367e8d8bef9SDimitry Andric void MemProfiler::instrumentMaskedLoadOrStore(const DataLayout &DL, Value *Mask, 368e8d8bef9SDimitry Andric Instruction *I, Value *Addr, 36904eeddc0SDimitry Andric Type *AccessTy, bool IsWrite) { 37004eeddc0SDimitry Andric auto *VTy = cast<FixedVectorType>(AccessTy); 371e8d8bef9SDimitry Andric uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType()); 372e8d8bef9SDimitry Andric unsigned Num = VTy->getNumElements(); 373e8d8bef9SDimitry Andric auto *Zero = ConstantInt::get(IntptrTy, 0); 374e8d8bef9SDimitry Andric for (unsigned Idx = 0; Idx < Num; ++Idx) { 375e8d8bef9SDimitry Andric Value *InstrumentedAddress = nullptr; 376e8d8bef9SDimitry Andric Instruction *InsertBefore = I; 377e8d8bef9SDimitry Andric if (auto *Vector = dyn_cast<ConstantVector>(Mask)) { 378e8d8bef9SDimitry Andric // dyn_cast as we might get UndefValue 379e8d8bef9SDimitry Andric if (auto *Masked = dyn_cast<ConstantInt>(Vector->getOperand(Idx))) { 380e8d8bef9SDimitry Andric if (Masked->isZero()) 381e8d8bef9SDimitry Andric // Mask is constant false, so no instrumentation needed. 382e8d8bef9SDimitry Andric continue; 383e8d8bef9SDimitry Andric // If we have a true or undef value, fall through to instrumentAddress. 384e8d8bef9SDimitry Andric // with InsertBefore == I 385e8d8bef9SDimitry Andric } 386e8d8bef9SDimitry Andric } else { 387e8d8bef9SDimitry Andric IRBuilder<> IRB(I); 388e8d8bef9SDimitry Andric Value *MaskElem = IRB.CreateExtractElement(Mask, Idx); 389e8d8bef9SDimitry Andric Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, I, false); 390e8d8bef9SDimitry Andric InsertBefore = ThenTerm; 391e8d8bef9SDimitry Andric } 392e8d8bef9SDimitry Andric 393e8d8bef9SDimitry Andric IRBuilder<> IRB(InsertBefore); 394e8d8bef9SDimitry Andric InstrumentedAddress = 395e8d8bef9SDimitry Andric IRB.CreateGEP(VTy, Addr, {Zero, ConstantInt::get(IntptrTy, Idx)}); 396e8d8bef9SDimitry Andric instrumentAddress(I, InsertBefore, InstrumentedAddress, ElemTypeSize, 397e8d8bef9SDimitry Andric IsWrite); 398e8d8bef9SDimitry Andric } 399e8d8bef9SDimitry Andric } 400e8d8bef9SDimitry Andric 401e8d8bef9SDimitry Andric void MemProfiler::instrumentMop(Instruction *I, const DataLayout &DL, 402e8d8bef9SDimitry Andric InterestingMemoryAccess &Access) { 403349cc55cSDimitry Andric // Skip instrumentation of stack accesses unless requested. 404349cc55cSDimitry Andric if (!ClStack && isa<AllocaInst>(getUnderlyingObject(Access.Addr))) { 405349cc55cSDimitry Andric if (Access.IsWrite) 406349cc55cSDimitry Andric ++NumSkippedStackWrites; 407349cc55cSDimitry Andric else 408349cc55cSDimitry Andric ++NumSkippedStackReads; 409349cc55cSDimitry Andric return; 410349cc55cSDimitry Andric } 411349cc55cSDimitry Andric 412e8d8bef9SDimitry Andric if (Access.IsWrite) 413e8d8bef9SDimitry Andric NumInstrumentedWrites++; 414e8d8bef9SDimitry Andric else 415e8d8bef9SDimitry Andric NumInstrumentedReads++; 416e8d8bef9SDimitry Andric 417e8d8bef9SDimitry Andric if (Access.MaybeMask) { 418e8d8bef9SDimitry Andric instrumentMaskedLoadOrStore(DL, Access.MaybeMask, I, Access.Addr, 41981ad6265SDimitry Andric Access.AccessTy, Access.IsWrite); 420e8d8bef9SDimitry Andric } else { 421e8d8bef9SDimitry Andric // Since the access counts will be accumulated across the entire allocation, 422e8d8bef9SDimitry Andric // we only update the shadow access count for the first location and thus 423e8d8bef9SDimitry Andric // don't need to worry about alignment and type size. 424e8d8bef9SDimitry Andric instrumentAddress(I, I, Access.Addr, Access.TypeSize, Access.IsWrite); 425e8d8bef9SDimitry Andric } 426e8d8bef9SDimitry Andric } 427e8d8bef9SDimitry Andric 428e8d8bef9SDimitry Andric void MemProfiler::instrumentAddress(Instruction *OrigIns, 429e8d8bef9SDimitry Andric Instruction *InsertBefore, Value *Addr, 430e8d8bef9SDimitry Andric uint32_t TypeSize, bool IsWrite) { 431e8d8bef9SDimitry Andric IRBuilder<> IRB(InsertBefore); 432e8d8bef9SDimitry Andric Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); 433e8d8bef9SDimitry Andric 434e8d8bef9SDimitry Andric if (ClUseCalls) { 435e8d8bef9SDimitry Andric IRB.CreateCall(MemProfMemoryAccessCallback[IsWrite], AddrLong); 436e8d8bef9SDimitry Andric return; 437e8d8bef9SDimitry Andric } 438e8d8bef9SDimitry Andric 439e8d8bef9SDimitry Andric // Create an inline sequence to compute shadow location, and increment the 440e8d8bef9SDimitry Andric // value by one. 441e8d8bef9SDimitry Andric Type *ShadowTy = Type::getInt64Ty(*C); 442e8d8bef9SDimitry Andric Type *ShadowPtrTy = PointerType::get(ShadowTy, 0); 443e8d8bef9SDimitry Andric Value *ShadowPtr = memToShadow(AddrLong, IRB); 444e8d8bef9SDimitry Andric Value *ShadowAddr = IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy); 445e8d8bef9SDimitry Andric Value *ShadowValue = IRB.CreateLoad(ShadowTy, ShadowAddr); 446e8d8bef9SDimitry Andric Value *Inc = ConstantInt::get(Type::getInt64Ty(*C), 1); 447e8d8bef9SDimitry Andric ShadowValue = IRB.CreateAdd(ShadowValue, Inc); 448e8d8bef9SDimitry Andric IRB.CreateStore(ShadowValue, ShadowAddr); 449e8d8bef9SDimitry Andric } 450e8d8bef9SDimitry Andric 451e8d8bef9SDimitry Andric // Create the variable for the profile file name. 452e8d8bef9SDimitry Andric void createProfileFileNameVar(Module &M) { 453e8d8bef9SDimitry Andric const MDString *MemProfFilename = 454e8d8bef9SDimitry Andric dyn_cast_or_null<MDString>(M.getModuleFlag("MemProfProfileFilename")); 455e8d8bef9SDimitry Andric if (!MemProfFilename) 456e8d8bef9SDimitry Andric return; 457e8d8bef9SDimitry Andric assert(!MemProfFilename->getString().empty() && 458e8d8bef9SDimitry Andric "Unexpected MemProfProfileFilename metadata with empty string"); 459e8d8bef9SDimitry Andric Constant *ProfileNameConst = ConstantDataArray::getString( 460e8d8bef9SDimitry Andric M.getContext(), MemProfFilename->getString(), true); 461e8d8bef9SDimitry Andric GlobalVariable *ProfileNameVar = new GlobalVariable( 462e8d8bef9SDimitry Andric M, ProfileNameConst->getType(), /*isConstant=*/true, 463e8d8bef9SDimitry Andric GlobalValue::WeakAnyLinkage, ProfileNameConst, MemProfFilenameVar); 464e8d8bef9SDimitry Andric Triple TT(M.getTargetTriple()); 465e8d8bef9SDimitry Andric if (TT.supportsCOMDAT()) { 466e8d8bef9SDimitry Andric ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage); 467e8d8bef9SDimitry Andric ProfileNameVar->setComdat(M.getOrInsertComdat(MemProfFilenameVar)); 468e8d8bef9SDimitry Andric } 469e8d8bef9SDimitry Andric } 470e8d8bef9SDimitry Andric 471e8d8bef9SDimitry Andric bool ModuleMemProfiler::instrumentModule(Module &M) { 472e8d8bef9SDimitry Andric // Create a module constructor. 473e8d8bef9SDimitry Andric std::string MemProfVersion = std::to_string(LLVM_MEM_PROFILER_VERSION); 474e8d8bef9SDimitry Andric std::string VersionCheckName = 475e8d8bef9SDimitry Andric ClInsertVersionCheck ? (MemProfVersionCheckNamePrefix + MemProfVersion) 476e8d8bef9SDimitry Andric : ""; 477e8d8bef9SDimitry Andric std::tie(MemProfCtorFunction, std::ignore) = 478e8d8bef9SDimitry Andric createSanitizerCtorAndInitFunctions(M, MemProfModuleCtorName, 479e8d8bef9SDimitry Andric MemProfInitName, /*InitArgTypes=*/{}, 480e8d8bef9SDimitry Andric /*InitArgs=*/{}, VersionCheckName); 481e8d8bef9SDimitry Andric 482e8d8bef9SDimitry Andric const uint64_t Priority = getCtorAndDtorPriority(TargetTriple); 483e8d8bef9SDimitry Andric appendToGlobalCtors(M, MemProfCtorFunction, Priority); 484e8d8bef9SDimitry Andric 485e8d8bef9SDimitry Andric createProfileFileNameVar(M); 486e8d8bef9SDimitry Andric 487e8d8bef9SDimitry Andric return true; 488e8d8bef9SDimitry Andric } 489e8d8bef9SDimitry Andric 490e8d8bef9SDimitry Andric void MemProfiler::initializeCallbacks(Module &M) { 491e8d8bef9SDimitry Andric IRBuilder<> IRB(*C); 492e8d8bef9SDimitry Andric 493e8d8bef9SDimitry Andric for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) { 494e8d8bef9SDimitry Andric const std::string TypeStr = AccessIsWrite ? "store" : "load"; 495e8d8bef9SDimitry Andric 496e8d8bef9SDimitry Andric SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy}; 497e8d8bef9SDimitry Andric SmallVector<Type *, 2> Args1{1, IntptrTy}; 498e8d8bef9SDimitry Andric MemProfMemoryAccessCallbackSized[AccessIsWrite] = 499e8d8bef9SDimitry Andric M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + TypeStr + "N", 500e8d8bef9SDimitry Andric FunctionType::get(IRB.getVoidTy(), Args2, false)); 501e8d8bef9SDimitry Andric 502e8d8bef9SDimitry Andric MemProfMemoryAccessCallback[AccessIsWrite] = 503e8d8bef9SDimitry Andric M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + TypeStr, 504e8d8bef9SDimitry Andric FunctionType::get(IRB.getVoidTy(), Args1, false)); 505e8d8bef9SDimitry Andric } 506e8d8bef9SDimitry Andric MemProfMemmove = M.getOrInsertFunction( 507e8d8bef9SDimitry Andric ClMemoryAccessCallbackPrefix + "memmove", IRB.getInt8PtrTy(), 508e8d8bef9SDimitry Andric IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy); 509e8d8bef9SDimitry Andric MemProfMemcpy = M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + "memcpy", 510e8d8bef9SDimitry Andric IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 511e8d8bef9SDimitry Andric IRB.getInt8PtrTy(), IntptrTy); 512e8d8bef9SDimitry Andric MemProfMemset = M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + "memset", 513e8d8bef9SDimitry Andric IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 514e8d8bef9SDimitry Andric IRB.getInt32Ty(), IntptrTy); 515e8d8bef9SDimitry Andric } 516e8d8bef9SDimitry Andric 517e8d8bef9SDimitry Andric bool MemProfiler::maybeInsertMemProfInitAtFunctionEntry(Function &F) { 518e8d8bef9SDimitry Andric // For each NSObject descendant having a +load method, this method is invoked 519e8d8bef9SDimitry Andric // by the ObjC runtime before any of the static constructors is called. 520e8d8bef9SDimitry Andric // Therefore we need to instrument such methods with a call to __memprof_init 521e8d8bef9SDimitry Andric // at the beginning in order to initialize our runtime before any access to 522e8d8bef9SDimitry Andric // the shadow memory. 523e8d8bef9SDimitry Andric // We cannot just ignore these methods, because they may call other 524e8d8bef9SDimitry Andric // instrumented functions. 525e8d8bef9SDimitry Andric if (F.getName().find(" load]") != std::string::npos) { 526e8d8bef9SDimitry Andric FunctionCallee MemProfInitFunction = 527e8d8bef9SDimitry Andric declareSanitizerInitFunction(*F.getParent(), MemProfInitName, {}); 528e8d8bef9SDimitry Andric IRBuilder<> IRB(&F.front(), F.front().begin()); 529e8d8bef9SDimitry Andric IRB.CreateCall(MemProfInitFunction, {}); 530e8d8bef9SDimitry Andric return true; 531e8d8bef9SDimitry Andric } 532e8d8bef9SDimitry Andric return false; 533e8d8bef9SDimitry Andric } 534e8d8bef9SDimitry Andric 535e8d8bef9SDimitry Andric bool MemProfiler::insertDynamicShadowAtFunctionEntry(Function &F) { 536e8d8bef9SDimitry Andric IRBuilder<> IRB(&F.front().front()); 537e8d8bef9SDimitry Andric Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal( 538e8d8bef9SDimitry Andric MemProfShadowMemoryDynamicAddress, IntptrTy); 539e8d8bef9SDimitry Andric if (F.getParent()->getPICLevel() == PICLevel::NotPIC) 540e8d8bef9SDimitry Andric cast<GlobalVariable>(GlobalDynamicAddress)->setDSOLocal(true); 541e8d8bef9SDimitry Andric DynamicShadowOffset = IRB.CreateLoad(IntptrTy, GlobalDynamicAddress); 542e8d8bef9SDimitry Andric return true; 543e8d8bef9SDimitry Andric } 544e8d8bef9SDimitry Andric 545e8d8bef9SDimitry Andric bool MemProfiler::instrumentFunction(Function &F) { 546e8d8bef9SDimitry Andric if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) 547e8d8bef9SDimitry Andric return false; 548e8d8bef9SDimitry Andric if (ClDebugFunc == F.getName()) 549e8d8bef9SDimitry Andric return false; 550e8d8bef9SDimitry Andric if (F.getName().startswith("__memprof_")) 551e8d8bef9SDimitry Andric return false; 552e8d8bef9SDimitry Andric 553e8d8bef9SDimitry Andric bool FunctionModified = false; 554e8d8bef9SDimitry Andric 555e8d8bef9SDimitry Andric // If needed, insert __memprof_init. 556e8d8bef9SDimitry Andric // This function needs to be called even if the function body is not 557e8d8bef9SDimitry Andric // instrumented. 558e8d8bef9SDimitry Andric if (maybeInsertMemProfInitAtFunctionEntry(F)) 559e8d8bef9SDimitry Andric FunctionModified = true; 560e8d8bef9SDimitry Andric 561e8d8bef9SDimitry Andric LLVM_DEBUG(dbgs() << "MEMPROF instrumenting:\n" << F << "\n"); 562e8d8bef9SDimitry Andric 563e8d8bef9SDimitry Andric initializeCallbacks(*F.getParent()); 564e8d8bef9SDimitry Andric 565e8d8bef9SDimitry Andric SmallVector<Instruction *, 16> ToInstrument; 566e8d8bef9SDimitry Andric 567e8d8bef9SDimitry Andric // Fill the set of memory operations to instrument. 568e8d8bef9SDimitry Andric for (auto &BB : F) { 569e8d8bef9SDimitry Andric for (auto &Inst : BB) { 570e8d8bef9SDimitry Andric if (isInterestingMemoryAccess(&Inst) || isa<MemIntrinsic>(Inst)) 571e8d8bef9SDimitry Andric ToInstrument.push_back(&Inst); 572e8d8bef9SDimitry Andric } 573e8d8bef9SDimitry Andric } 574e8d8bef9SDimitry Andric 57581ad6265SDimitry Andric if (ToInstrument.empty()) { 57681ad6265SDimitry Andric LLVM_DEBUG(dbgs() << "MEMPROF done instrumenting: " << FunctionModified 57781ad6265SDimitry Andric << " " << F << "\n"); 57881ad6265SDimitry Andric 57981ad6265SDimitry Andric return FunctionModified; 58081ad6265SDimitry Andric } 58181ad6265SDimitry Andric 58281ad6265SDimitry Andric FunctionModified |= insertDynamicShadowAtFunctionEntry(F); 58381ad6265SDimitry Andric 584e8d8bef9SDimitry Andric int NumInstrumented = 0; 585e8d8bef9SDimitry Andric for (auto *Inst : ToInstrument) { 586e8d8bef9SDimitry Andric if (ClDebugMin < 0 || ClDebugMax < 0 || 587e8d8bef9SDimitry Andric (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) { 588*bdd1243dSDimitry Andric std::optional<InterestingMemoryAccess> Access = 589e8d8bef9SDimitry Andric isInterestingMemoryAccess(Inst); 590e8d8bef9SDimitry Andric if (Access) 591e8d8bef9SDimitry Andric instrumentMop(Inst, F.getParent()->getDataLayout(), *Access); 592e8d8bef9SDimitry Andric else 593e8d8bef9SDimitry Andric instrumentMemIntrinsic(cast<MemIntrinsic>(Inst)); 594e8d8bef9SDimitry Andric } 595e8d8bef9SDimitry Andric NumInstrumented++; 596e8d8bef9SDimitry Andric } 597e8d8bef9SDimitry Andric 598e8d8bef9SDimitry Andric if (NumInstrumented > 0) 599e8d8bef9SDimitry Andric FunctionModified = true; 600e8d8bef9SDimitry Andric 601e8d8bef9SDimitry Andric LLVM_DEBUG(dbgs() << "MEMPROF done instrumenting: " << FunctionModified << " " 602e8d8bef9SDimitry Andric << F << "\n"); 603e8d8bef9SDimitry Andric 604e8d8bef9SDimitry Andric return FunctionModified; 605e8d8bef9SDimitry Andric } 606