CDSPass.cpp

   1 //===-- CDSPass.cpp - xxx -------------------------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 // This file is distributed under the University of Illinois Open Source
   7 // License. See LICENSE.TXT for details.
   8 //
   9 //===----------------------------------------------------------------------===//
  10 //
  11 // This file is a modified version of ThreadSanitizer.cpp, a part of a race detector.
  12 //
  13 // The tool is under development, for the details about previous versions see
  14 // http://code.google.com/p/data-race-test
  15 //
  16 // The instrumentation phase is quite simple:
  17 //   - Insert calls to run-time library before every memory access.
  18 //      - Optimizations may apply to avoid instrumenting some of the accesses.
  19 //   - Insert calls at function entry/exit.
  20 // The rest is handled by the run-time library.
  21 //===----------------------------------------------------------------------===//
  22
  23 #include "llvm/ADT/Statistic.h"
  24 #include "llvm/ADT/StringExtras.h"
  25 #include "llvm/ADT/SmallString.h"
  26 #include "llvm/Analysis/ValueTracking.h"
  27 #include "llvm/Analysis/CaptureTracking.h"
  28 #include "llvm/IR/BasicBlock.h"
  29 #include "llvm/IR/Function.h"
  30 #include "llvm/IR/IRBuilder.h"
  31 #include "llvm/IR/Instructions.h"
  32 #include "llvm/IR/LLVMContext.h"
  33 #include "llvm/IR/LegacyPassManager.h"
  34 #include "llvm/IR/Module.h"
  35 #include "llvm/IR/PassManager.h"
  36 #include "llvm/Pass.h"
  37 #include "llvm/ProfileData/InstrProf.h"
  38 #include "llvm/Support/raw_ostream.h"
  39 #include "llvm/Support/AtomicOrdering.h"
  40 #include "llvm/Support/Debug.h"
  41 #include "llvm/Transforms/Scalar.h"
  42 #include "llvm/Transforms/Utils/Local.h"
  43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  44 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
  45 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
  46 #include <vector>
  47
  48 using namespace llvm;
  49
  50 #define DEBUG_TYPE "CDS"
  51 #include <llvm/IR/DebugLoc.h>
  52
  53 Value *getPosition( Instruction * I, IRBuilder <> IRB, bool print = false)
  54 {
  55         const DebugLoc & debug_location = I->getDebugLoc ();
  56         std::string position_string;
  57         {
  58                 llvm::raw_string_ostream position_stream (position_string);
  59                 debug_location . print (position_stream);
  60         }
  61
  62         if (print) {
  63                 errs() << position_string << "\n";
  64         }
  65
  66         return IRB.CreateGlobalStringPtr (position_string);
  67 }
  68
  69 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
  70 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
  71 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
  72 // STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
  73 // STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
  74
  75 STATISTIC(NumOmittedReadsBeforeWrite,
  76           "Number of reads ignored due to following writes");
  77 STATISTIC(NumOmittedReadsFromConstantGlobals,
  78           "Number of reads from constant globals");
  79 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
  80 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
  81
  82 Type * OrdTy;
  83
  84 Type * Int8PtrTy;
  85 Type * Int16PtrTy;
  86 Type * Int32PtrTy;
  87 Type * Int64PtrTy;
  88
  89 Type * VoidTy;
  90
  91 static const size_t kNumberOfAccessSizes = 4;
  92 static const int volatile_order = 6;
  93
  94 int getAtomicOrderIndex(AtomicOrdering order){
  95         switch (order) {
  96                 case AtomicOrdering::Monotonic:
  97                         return (int)AtomicOrderingCABI::relaxed;
  98                 //  case AtomicOrdering::Consume:         // not specified yet
  99                 //    return AtomicOrderingCABI::consume;
 100                 case AtomicOrdering::Acquire:
 101                         return (int)AtomicOrderingCABI::acquire;
 102                 case AtomicOrdering::Release:
 103                         return (int)AtomicOrderingCABI::release;
 104                 case AtomicOrdering::AcquireRelease:
 105                         return (int)AtomicOrderingCABI::acq_rel;
 106                 case AtomicOrdering::SequentiallyConsistent:
 107                         return (int)AtomicOrderingCABI::seq_cst;
 108                 default:
 109                         // unordered or Not Atomic
 110                         return -1;
 111         }
 112 }
 113
 114 namespace {
 115         struct CDSPass : public FunctionPass {
 116                 static char ID;
 117                 CDSPass() : FunctionPass(ID) {}
 118                 bool runOnFunction(Function &F) override;
 119
 120         private:
 121                 void initializeCallbacks(Module &M);
 122                 bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
 123                 bool instrumentVolatile(Instruction *I, const DataLayout &DL);
 124                 bool isAtomicCall(Instruction *I);
 125                 bool instrumentAtomic(Instruction *I, const DataLayout &DL);
 126                 bool instrumentAtomicCall(CallInst *CI, const DataLayout &DL);
 127                 void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
 128                                                                                         SmallVectorImpl<Instruction *> &All,
 129                                                                                         const DataLayout &DL);
 130                 bool addrPointsToConstantData(Value *Addr);
 131                 int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
 132
 133                 // Callbacks to run-time library are computed in doInitialization.
 134                 Constant * CDSFuncEntry;
 135                 Constant * CDSFuncExit;
 136
 137                 Constant * CDSLoad[kNumberOfAccessSizes];
 138                 Constant * CDSStore[kNumberOfAccessSizes];
 139                 Constant * CDSVolatileLoad[kNumberOfAccessSizes];
 140                 Constant * CDSVolatileStore[kNumberOfAccessSizes];
 141                 Constant * CDSAtomicInit[kNumberOfAccessSizes];
 142                 Constant * CDSAtomicLoad[kNumberOfAccessSizes];
 143                 Constant * CDSAtomicStore[kNumberOfAccessSizes];
 144                 Constant * CDSAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
 145                 Constant * CDSAtomicCAS_V1[kNumberOfAccessSizes];
 146                 Constant * CDSAtomicCAS_V2[kNumberOfAccessSizes];
 147                 Constant * CDSAtomicThreadFence;
 148
 149                 std::vector<StringRef> AtomicFuncNames;
 150                 std::vector<StringRef> PartialAtomicFuncNames;
 151         };
 152 }
 153
 154 static bool isVtableAccess(Instruction *I) {
 155         if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 156                 return Tag->isTBAAVtableAccess();
 157         return false;
 158 }
 159
 160 void CDSPass::initializeCallbacks(Module &M) {
 161         LLVMContext &Ctx = M.getContext();
 162
 163         Type * Int1Ty = Type::getInt1Ty(Ctx);
 164         OrdTy = Type::getInt32Ty(Ctx);
 165
 166         Int8PtrTy  = Type::getInt8PtrTy(Ctx);
 167         Int16PtrTy = Type::getInt16PtrTy(Ctx);
 168         Int32PtrTy = Type::getInt32PtrTy(Ctx);
 169         Int64PtrTy = Type::getInt64PtrTy(Ctx);
 170
 171         VoidTy = Type::getVoidTy(Ctx);
 172
 173         CDSFuncEntry = M.getOrInsertFunction("cds_func_entry",
 174                                                                 VoidTy, Int8PtrTy);
 175         CDSFuncExit = M.getOrInsertFunction("cds_func_exit",
 176                                                                 VoidTy, Int8PtrTy);
 177
 178         // Get the function to call from our untime library.
 179         for (unsigned i = 0; i < kNumberOfAccessSizes; i++) {
 180                 const unsigned ByteSize = 1U << i;
 181                 const unsigned BitSize = ByteSize * 8;
 182
 183                 std::string ByteSizeStr = utostr(ByteSize);
 184                 std::string BitSizeStr = utostr(BitSize);
 185
 186                 Type *Ty = Type::getIntNTy(Ctx, BitSize);
 187                 Type *PtrTy = Ty->getPointerTo();
 188
 189                 // uint8_t cds_atomic_load8 (void * obj, int atomic_index)
 190                 // void cds_atomic_store8 (void * obj, int atomic_index, uint8_t val)
 191                 SmallString<32> LoadName("cds_load" + BitSizeStr);
 192                 SmallString<32> StoreName("cds_store" + BitSizeStr);
 193                 SmallString<32> VolatileLoadName("cds_volatile_load" + BitSizeStr);
 194                 SmallString<32> VolatileStoreName("cds_volatile_store" + BitSizeStr);
 195                 SmallString<32> AtomicInitName("cds_atomic_init" + BitSizeStr);
 196                 SmallString<32> AtomicLoadName("cds_atomic_load" + BitSizeStr);
 197                 SmallString<32> AtomicStoreName("cds_atomic_store" + BitSizeStr);
 198
 199                 CDSLoad[i]  = M.getOrInsertFunction(LoadName, VoidTy, PtrTy);
 200                 CDSStore[i] = M.getOrInsertFunction(StoreName, VoidTy, PtrTy);
 201                 CDSVolatileLoad[i]  = M.getOrInsertFunction(VolatileLoadName,
 202                                                                         Ty, PtrTy, OrdTy, Int8PtrTy);
 203                 CDSVolatileStore[i] = M.getOrInsertFunction(VolatileStoreName,
 204                                                                         VoidTy, PtrTy, Ty, OrdTy, Int8PtrTy);
 205                 CDSAtomicInit[i] = M.getOrInsertFunction(AtomicInitName,
 206                                                                 VoidTy, PtrTy, Ty, Int8PtrTy);
 207                 CDSAtomicLoad[i]  = M.getOrInsertFunction(AtomicLoadName,
 208                                                                 Ty, PtrTy, OrdTy, Int8PtrTy);
 209                 CDSAtomicStore[i] = M.getOrInsertFunction(AtomicStoreName,
 210                                                                 VoidTy, PtrTy, Ty, OrdTy, Int8PtrTy);
 211
 212                 for (int op = AtomicRMWInst::FIRST_BINOP;
 213                         op <= AtomicRMWInst::LAST_BINOP; ++op) {
 214                         CDSAtomicRMW[op][i] = nullptr;
 215                         std::string NamePart;
 216
 217                         if (op == AtomicRMWInst::Xchg)
 218                                 NamePart = "_exchange";
 219                         else if (op == AtomicRMWInst::Add)
 220                                 NamePart = "_fetch_add";
 221                         else if (op == AtomicRMWInst::Sub)
 222                                 NamePart = "_fetch_sub";
 223                         else if (op == AtomicRMWInst::And)
 224                                 NamePart = "_fetch_and";
 225                         else if (op == AtomicRMWInst::Or)
 226                                 NamePart = "_fetch_or";
 227                         else if (op == AtomicRMWInst::Xor)
 228                                 NamePart = "_fetch_xor";
 229                         else
 230                                 continue;
 231
 232                         SmallString<32> AtomicRMWName("cds_atomic" + NamePart + BitSizeStr);
 233                         CDSAtomicRMW[op][i] = M.getOrInsertFunction(AtomicRMWName,
 234                                                                                 Ty, PtrTy, Ty, OrdTy, Int8PtrTy);
 235                 }
 236
 237                 // only supportes strong version
 238                 SmallString<32> AtomicCASName_V1("cds_atomic_compare_exchange" + BitSizeStr + "_v1");
 239                 SmallString<32> AtomicCASName_V2("cds_atomic_compare_exchange" + BitSizeStr + "_v2");
 240                 CDSAtomicCAS_V1[i] = M.getOrInsertFunction(AtomicCASName_V1,
 241                                                                 Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, Int8PtrTy);
 242                 CDSAtomicCAS_V2[i] = M.getOrInsertFunction(AtomicCASName_V2,
 243                                                                 Int1Ty, PtrTy, PtrTy, Ty, OrdTy, OrdTy, Int8PtrTy);
 244         }
 245
 246         CDSAtomicThreadFence = M.getOrInsertFunction("cds_atomic_thread_fence",
 247                                                                                                         VoidTy, OrdTy, Int8PtrTy);
 248 }
 249
 250 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
 251         // Peel off GEPs and BitCasts.
 252         Addr = Addr->stripInBoundsOffsets();
 253
 254         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 255                 if (GV->hasSection()) {
 256                         StringRef SectionName = GV->getSection();
 257                         // Check if the global is in the PGO counters section.
 258                         auto OF = Triple(M->getTargetTriple()).getObjectFormat();
 259                         if (SectionName.endswith(
 260                               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
 261                                 return false;
 262                 }
 263
 264                 // Check if the global is private gcov data.
 265                 if (GV->getName().startswith("__llvm_gcov") ||
 266                 GV->getName().startswith("__llvm_gcda"))
 267                 return false;
 268         }
 269
 270         // Do not instrument acesses from different address spaces; we cannot deal
 271         // with them.
 272         if (Addr) {
 273                 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
 274                 if (PtrTy->getPointerAddressSpace() != 0)
 275                         return false;
 276         }
 277
 278         return true;
 279 }
 280
 281 bool CDSPass::addrPointsToConstantData(Value *Addr) {
 282         // If this is a GEP, just analyze its pointer operand.
 283         if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
 284                 Addr = GEP->getPointerOperand();
 285
 286         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 287                 if (GV->isConstant()) {
 288                         // Reads from constant globals can not race with any writes.
 289                         NumOmittedReadsFromConstantGlobals++;
 290                         return true;
 291                 }
 292         } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
 293                 if (isVtableAccess(L)) {
 294                         // Reads from a vtable pointer can not race with any writes.
 295                         NumOmittedReadsFromVtable++;
 296                         return true;
 297                 }
 298         }
 299         return false;
 300 }
 301
 302 bool CDSPass::runOnFunction(Function &F) {
 303         if (F.getName() == "main") {
 304                 F.setName("user_main");
 305                 errs() << "main replaced by user_main\n";
 306         }
 307
 308         if (true) {
 309                 initializeCallbacks( *F.getParent() );
 310
 311                 AtomicFuncNames =
 312                 {
 313                         "atomic_init", "atomic_load", "atomic_store",
 314                         "atomic_fetch_", "atomic_exchange", "atomic_compare_exchange_"
 315                 };
 316
 317                 PartialAtomicFuncNames =
 318                 {
 319                         "load", "store", "fetch", "exchange", "compare_exchange_"
 320                 };
 321
 322                 SmallVector<Instruction*, 8> AllLoadsAndStores;
 323                 SmallVector<Instruction*, 8> LocalLoadsAndStores;
 324                 SmallVector<Instruction*, 8> VolatileLoadsAndStores;
 325                 SmallVector<Instruction*, 8> AtomicAccesses;
 326
 327                 std::vector<Instruction *> worklist;
 328
 329                 bool Res = false;
 330                 bool HasAtomic = false;
 331                 const DataLayout &DL = F.getParent()->getDataLayout();
 332
 333                 // errs() << "--- " << F.getName() << "---\n";
 334
 335                 for (auto &B : F) {
 336                         for (auto &I : B) {
 337                                 if ( (&I)->isAtomic() || isAtomicCall(&I) ) {
 338                                         AtomicAccesses.push_back(&I);
 339                                         HasAtomic = true;
 340                                 } else if (isa<LoadInst>(I) || isa<StoreInst>(I)) {
 341                                         LoadInst *LI = dyn_cast<LoadInst>(&I);
 342                                         StoreInst *SI = dyn_cast<StoreInst>(&I);
 343                                         bool isVolatile = ( LI ? LI->isVolatile() : SI->isVolatile() );
 344
 345                                         if (isVolatile)
 346                                                 VolatileLoadsAndStores.push_back(&I);
 347                                         else
 348                                                 LocalLoadsAndStores.push_back(&I);
 349                                 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
 350                                         // not implemented yet
 351                                 }
 352                         }
 353
 354                         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
 355                 }
 356
 357                 for (auto Inst : AllLoadsAndStores) {
 358                         Res |= instrumentLoadOrStore(Inst, DL);
 359                 }
 360
 361                 for (auto Inst : VolatileLoadsAndStores) {
 362                         Res |= instrumentVolatile(Inst, DL);
 363                 }
 364
 365                 for (auto Inst : AtomicAccesses) {
 366                         Res |= instrumentAtomic(Inst, DL);
 367                 }
 368
 369                 // only instrument functions that contain atomics
 370                 if (Res && HasAtomic) {
 371                         IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 372                         /* Unused for now
 373                         Value *ReturnAddress = IRB.CreateCall(
 374                                 Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 375                                 IRB.getInt32(0));
 376                         */
 377
 378                         Value * FuncName = IRB.CreateGlobalStringPtr(F.getName());
 379                         IRB.CreateCall(CDSFuncEntry, FuncName);
 380
 381                         EscapeEnumerator EE(F, "cds_cleanup", true);
 382                         while (IRBuilder<> *AtExit = EE.Next()) {
 383                           AtExit->CreateCall(CDSFuncExit, FuncName);
 384                         }
 385
 386                         Res = true;
 387                 }
 388         }
 389
 390         return false;
 391 }
 392
 393 void CDSPass::chooseInstructionsToInstrument(
 394         SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
 395         const DataLayout &DL) {
 396         SmallPtrSet<Value*, 8> WriteTargets;
 397         // Iterate from the end.
 398         for (Instruction *I : reverse(Local)) {
 399                 if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
 400                         Value *Addr = Store->getPointerOperand();
 401                         if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
 402                                 continue;
 403                         WriteTargets.insert(Addr);
 404                 } else {
 405                         LoadInst *Load = cast<LoadInst>(I);
 406                         Value *Addr = Load->getPointerOperand();
 407                         if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
 408                                 continue;
 409                         if (WriteTargets.count(Addr)) {
 410                                 // We will write to this temp, so no reason to analyze the read.
 411                                 NumOmittedReadsBeforeWrite++;
 412                                 continue;
 413                         }
 414                         if (addrPointsToConstantData(Addr)) {
 415                                 // Addr points to some constant data -- it can not race with any writes.
 416                                 continue;
 417                         }
 418                 }
 419                 Value *Addr = isa<StoreInst>(*I)
 420                         ? cast<StoreInst>(I)->getPointerOperand()
 421                         : cast<LoadInst>(I)->getPointerOperand();
 422                 if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
 423                                 !PointerMayBeCaptured(Addr, true, true)) {
 424                         // The variable is addressable but not captured, so it cannot be
 425                         // referenced from a different thread and participate in a data race
 426                         // (see llvm/Analysis/CaptureTracking.h for details).
 427                         NumOmittedNonCaptured++;
 428                         continue;
 429                 }
 430                 All.push_back(I);
 431         }
 432         Local.clear();
 433 }
 434
 435
 436 bool CDSPass::instrumentLoadOrStore(Instruction *I,
 437                                                                         const DataLayout &DL) {
 438         IRBuilder<> IRB(I);
 439         bool IsWrite = isa<StoreInst>(*I);
 440         Value *Addr = IsWrite
 441                 ? cast<StoreInst>(I)->getPointerOperand()
 442                 : cast<LoadInst>(I)->getPointerOperand();
 443
 444         // swifterror memory addresses are mem2reg promoted by instruction selection.
 445         // As such they cannot have regular uses like an instrumentation function and
 446         // it makes no sense to track them as memory.
 447         if (Addr->isSwiftError())
 448         return false;
 449
 450         int Idx = getMemoryAccessFuncIndex(Addr, DL);
 451         if (Idx < 0)
 452                 return false;
 453
 454 //  not supported by CDS yet
 455 /*  if (IsWrite && isVtableAccess(I)) {
 456     LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
 457     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
 458     // StoredValue may be a vector type if we are storing several vptrs at once.
 459     // In this case, just take the first element of the vector since this is
 460     // enough to find vptr races.
 461     if (isa<VectorType>(StoredValue->getType()))
 462       StoredValue = IRB.CreateExtractElement(
 463           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 464     if (StoredValue->getType()->isIntegerTy())
 465       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 466     // Call TsanVptrUpdate.
 467     IRB.CreateCall(TsanVptrUpdate,
 468                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 469                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
 470     NumInstrumentedVtableWrites++;
 471     return true;
 472   }
 473
 474   if (!IsWrite && isVtableAccess(I)) {
 475     IRB.CreateCall(TsanVptrLoad,
 476                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 477     NumInstrumentedVtableReads++;
 478     return true;
 479   }
 480 */
 481
 482         Value *OnAccessFunc = nullptr;
 483         OnAccessFunc = IsWrite ? CDSStore[Idx] : CDSLoad[Idx];
 484
 485         Type *ArgType = IRB.CreatePointerCast(Addr, Addr->getType())->getType();
 486
 487         if ( ArgType != Int8PtrTy && ArgType != Int16PtrTy &&
 488                         ArgType != Int32PtrTy && ArgType != Int64PtrTy ) {
 489                 // if other types of load or stores are passed in
 490                 return false;
 491         }
 492         IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, Addr->getType()));
 493         if (IsWrite) NumInstrumentedWrites++;
 494         else         NumInstrumentedReads++;
 495         return true;
 496 }
 497
 498 bool CDSPass::instrumentVolatile(Instruction * I, const DataLayout &DL) {
 499         IRBuilder<> IRB(I);
 500         Value *position = getPosition(I, IRB);
 501
 502         if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 503                 assert( LI->isVolatile() );
 504                 Value *Addr = LI->getPointerOperand();
 505                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 506                 if (Idx < 0)
 507                         return false;
 508
 509                 Value *order = ConstantInt::get(OrdTy, volatile_order);
 510                 Value *args[] = {Addr, order, position};
 511                 Instruction* funcInst=CallInst::Create(CDSVolatileLoad[Idx], args);
 512                 ReplaceInstWithInst(LI, funcInst);
 513         } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 514                 assert( SI->isVolatile() );
 515                 Value *Addr = SI->getPointerOperand();
 516                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 517                 if (Idx < 0)
 518                         return false;
 519
 520                 Value *val = SI->getValueOperand();
 521                 Value *order = ConstantInt::get(OrdTy, volatile_order);
 522                 Value *args[] = {Addr, val, order, position};
 523                 Instruction* funcInst=CallInst::Create(CDSVolatileStore[Idx], args);
 524                 ReplaceInstWithInst(SI, funcInst);
 525         } else {
 526                 return false;
 527         }
 528
 529         return true;
 530 }
 531
 532 bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
 533         IRBuilder<> IRB(I);
 534
 535         if (auto *CI = dyn_cast<CallInst>(I)) {
 536                 return instrumentAtomicCall(CI, DL);
 537         }
 538
 539         Value *position = getPosition(I, IRB);
 540
 541         if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 542                 Value *Addr = LI->getPointerOperand();
 543                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 544                 if (Idx < 0)
 545                         return false;
 546
 547                 int atomic_order_index = getAtomicOrderIndex(LI->getOrdering());
 548                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
 549                 Value *args[] = {Addr, order, position};
 550                 Instruction* funcInst=CallInst::Create(CDSAtomicLoad[Idx], args);
 551                 ReplaceInstWithInst(LI, funcInst);
 552         } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 553                 Value *Addr = SI->getPointerOperand();
 554                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 555                 if (Idx < 0)
 556                         return false;
 557
 558                 int atomic_order_index = getAtomicOrderIndex(SI->getOrdering());
 559                 Value *val = SI->getValueOperand();
 560                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
 561                 Value *args[] = {Addr, val, order, position};
 562                 Instruction* funcInst=CallInst::Create(CDSAtomicStore[Idx], args);
 563                 ReplaceInstWithInst(SI, funcInst);
 564         } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 565                 Value *Addr = RMWI->getPointerOperand();
 566                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 567                 if (Idx < 0)
 568                         return false;
 569
 570                 int atomic_order_index = getAtomicOrderIndex(RMWI->getOrdering());
 571                 Value *val = RMWI->getValOperand();
 572                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
 573                 Value *args[] = {Addr, val, order, position};
 574                 Instruction* funcInst = CallInst::Create(CDSAtomicRMW[RMWI->getOperation()][Idx], args);
 575                 ReplaceInstWithInst(RMWI, funcInst);
 576         } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 577                 IRBuilder<> IRB(CASI);
 578
 579                 Value *Addr = CASI->getPointerOperand();
 580                 int Idx=getMemoryAccessFuncIndex(Addr, DL);
 581                 if (Idx < 0)
 582                         return false;
 583
 584                 const unsigned ByteSize = 1U << Idx;
 585                 const unsigned BitSize = ByteSize * 8;
 586                 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 587                 Type *PtrTy = Ty->getPointerTo();
 588
 589                 Value *CmpOperand = IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
 590                 Value *NewOperand = IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
 591
 592                 int atomic_order_index_succ = getAtomicOrderIndex(CASI->getSuccessOrdering());
 593                 int atomic_order_index_fail = getAtomicOrderIndex(CASI->getFailureOrdering());
 594                 Value *order_succ = ConstantInt::get(OrdTy, atomic_order_index_succ);
 595                 Value *order_fail = ConstantInt::get(OrdTy, atomic_order_index_fail);
 596
 597                 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 598                                                  CmpOperand, NewOperand,
 599                                                  order_succ, order_fail, position};
 600
 601                 CallInst *funcInst = IRB.CreateCall(CDSAtomicCAS_V1[Idx], Args);
 602                 Value *Success = IRB.CreateICmpEQ(funcInst, CmpOperand);
 603
 604                 Value *OldVal = funcInst;
 605                 Type *OrigOldValTy = CASI->getNewValOperand()->getType();
 606                 if (Ty != OrigOldValTy) {
 607                         // The value is a pointer, so we need to cast the return value.
 608                         OldVal = IRB.CreateIntToPtr(funcInst, OrigOldValTy);
 609                 }
 610
 611                 Value *Res =
 612                   IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
 613                 Res = IRB.CreateInsertValue(Res, Success, 1);
 614
 615                 I->replaceAllUsesWith(Res);
 616                 I->eraseFromParent();
 617         } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 618                 int atomic_order_index = getAtomicOrderIndex(FI->getOrdering());
 619                 Value *order = ConstantInt::get(OrdTy, atomic_order_index);
 620                 Value *Args[] = {order, position};
 621
 622                 CallInst *funcInst = CallInst::Create(CDSAtomicThreadFence, Args);
 623                 ReplaceInstWithInst(FI, funcInst);
 624                 // errs() << "Thread Fences replaced\n";
 625         }
 626         return true;
 627 }
 628
 629 bool CDSPass::isAtomicCall(Instruction *I) {
 630         if ( auto *CI = dyn_cast<CallInst>(I) ) {
 631                 Function *fun = CI->getCalledFunction();
 632                 if (fun == NULL)
 633                         return false;
 634
 635                 StringRef funName = fun->getName();
 636
 637                 // todo: come up with better rules for function name checking
 638                 for (StringRef name : AtomicFuncNames) {
 639                         if ( funName.contains(name) )
 640                                 return true;
 641                 }
 642
 643                 for (StringRef PartialName : PartialAtomicFuncNames) {
 644                         if (funName.contains(PartialName) &&
 645                                         funName.contains("atomic") )
 646                                 return true;
 647                 }
 648         }
 649
 650         return false;
 651 }
 652
 653 bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
 654         IRBuilder<> IRB(CI);
 655         Function *fun = CI->getCalledFunction();
 656         StringRef funName = fun->getName();
 657         std::vector<Value *> parameters;
 658
 659         User::op_iterator begin = CI->arg_begin();
 660         User::op_iterator end = CI->arg_end();
 661         for (User::op_iterator it = begin; it != end; ++it) {
 662                 Value *param = *it;
 663                 parameters.push_back(param);
 664         }
 665
 666         // obtain source line number of the CallInst
 667         Value *position = getPosition(CI, IRB);
 668
 669         // the pointer to the address is always the first argument
 670         Value *OrigPtr = parameters[0];
 671
 672         int Idx = getMemoryAccessFuncIndex(OrigPtr, DL);
 673         if (Idx < 0)
 674                 return false;
 675
 676         const unsigned ByteSize = 1U << Idx;
 677         const unsigned BitSize = ByteSize * 8;
 678         Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 679         Type *PtrTy = Ty->getPointerTo();
 680
 681         // atomic_init; args = {obj, order}
 682         if (funName.contains("atomic_init")) {
 683                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 684                 Value *val = IRB.CreateBitOrPointerCast(parameters[1], Ty);
 685                 Value *args[] = {ptr, val, position};
 686
 687                 Instruction* funcInst = CallInst::Create(CDSAtomicInit[Idx], args);
 688                 ReplaceInstWithInst(CI, funcInst);
 689
 690                 return true;
 691         }
 692
 693         // atomic_load; args = {obj, order}
 694         if (funName.contains("atomic_load")) {
 695                 bool isExplicit = funName.contains("atomic_load_explicit");
 696
 697                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 698                 Value *order;
 699                 if (isExplicit)
 700                         order = IRB.CreateBitOrPointerCast(parameters[1], OrdTy);
 701                 else
 702                         order = ConstantInt::get(OrdTy,
 703                                                         (int) AtomicOrderingCABI::seq_cst);
 704                 Value *args[] = {ptr, order, position};
 705
 706                 Instruction* funcInst = CallInst::Create(CDSAtomicLoad[Idx], args);
 707                 ReplaceInstWithInst(CI, funcInst);
 708
 709                 return true;
 710         } else if (funName.contains("atomic") &&
 711                                         funName.contains("load") ) {
 712                 // does this version of call always have an atomic order as an argument?
 713                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 714                 Value *order = IRB.CreateBitOrPointerCast(parameters[1], OrdTy);
 715                 Value *args[] = {ptr, order, position};
 716
 717                 if (!CI->getType()->isPointerTy()) {
 718                         return false;
 719                 }
 720
 721                 CallInst *funcInst = IRB.CreateCall(CDSAtomicLoad[Idx], args);
 722                 Value *RetVal = IRB.CreateIntToPtr(funcInst, CI->getType());
 723
 724                 CI->replaceAllUsesWith(RetVal);
 725                 CI->eraseFromParent();
 726
 727                 return true;
 728         }
 729
 730         // atomic_store; args = {obj, val, order}
 731         if (funName.contains("atomic_store")) {
 732                 bool isExplicit = funName.contains("atomic_store_explicit");
 733                 Value *OrigVal = parameters[1];
 734
 735                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 736                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
 737                 Value *order;
 738                 if (isExplicit)
 739                         order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
 740                 else
 741                         order = ConstantInt::get(OrdTy,
 742                                                         (int) AtomicOrderingCABI::seq_cst);
 743                 Value *args[] = {ptr, val, order, position};
 744
 745                 Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
 746                 ReplaceInstWithInst(CI, funcInst);
 747
 748                 return true;
 749         } else if (funName.contains("atomic") &&
 750                                         funName.contains("EEEE5store") ) {
 751                 // does this version of call always have an atomic order as an argument?
 752                 Value *OrigVal = parameters[1];
 753
 754                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 755                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
 756                 Value *order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
 757                 Value *args[] = {ptr, val, order, position};
 758
 759                 Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
 760                 ReplaceInstWithInst(CI, funcInst);
 761
 762                 return true;
 763         }
 764
 765         // atomic_fetch_*; args = {obj, val, order}
 766         if (funName.contains("atomic_fetch_") ||
 767                         funName.contains("atomic_exchange") ) {
 768                 bool isExplicit = funName.contains("_explicit");
 769                 Value *OrigVal = parameters[1];
 770
 771                 int op;
 772                 if ( funName.contains("_fetch_add") )
 773                         op = AtomicRMWInst::Add;
 774                 else if ( funName.contains("_fetch_sub") )
 775                         op = AtomicRMWInst::Sub;
 776                 else if ( funName.contains("_fetch_and") )
 777                         op = AtomicRMWInst::And;
 778                 else if ( funName.contains("_fetch_or") )
 779                         op = AtomicRMWInst::Or;
 780                 else if ( funName.contains("_fetch_xor") )
 781                         op = AtomicRMWInst::Xor;
 782                 else if ( funName.contains("atomic_exchange") )
 783                         op = AtomicRMWInst::Xchg;
 784                 else {
 785                         errs() << "Unknown atomic read-modify-write operation\n";
 786                         return false;
 787                 }
 788
 789                 Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 790                 Value *val = IRB.CreatePointerCast(OrigVal, Ty);
 791                 Value *order;
 792                 if (isExplicit)
 793                         order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
 794                 else
 795                         order = ConstantInt::get(OrdTy,
 796                                                         (int) AtomicOrderingCABI::seq_cst);
 797                 Value *args[] = {ptr, val, order, position};
 798
 799                 Instruction* funcInst = CallInst::Create(CDSAtomicRMW[op][Idx], args);
 800                 ReplaceInstWithInst(CI, funcInst);
 801
 802                 return true;
 803         } else if (funName.contains("fetch")) {
 804                 errs() << "atomic exchange captured. Not implemented yet. ";
 805                 errs() << "See source file :";
 806                 getPosition(CI, IRB, true);
 807         } else if (funName.contains("exchange") &&
 808                         !funName.contains("compare_exchange") ) {
 809                 errs() << "atomic exchange captured. Not implemented yet. ";
 810                 errs() << "See source file :";
 811                 getPosition(CI, IRB, true);
 812         }
 813
 814         /* atomic_compare_exchange_*;
 815            args = {obj, expected, new value, order1, order2}
 816         */
 817         if ( funName.contains("atomic_compare_exchange_") ) {
 818                 bool isExplicit = funName.contains("_explicit");
 819
 820                 Value *Addr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 821                 Value *CmpOperand = IRB.CreatePointerCast(parameters[1], PtrTy);
 822                 Value *NewOperand = IRB.CreateBitOrPointerCast(parameters[2], Ty);
 823
 824                 Value *order_succ, *order_fail;
 825                 if (isExplicit) {
 826                         order_succ = IRB.CreateBitOrPointerCast(parameters[3], OrdTy);
 827                         order_fail = IRB.CreateBitOrPointerCast(parameters[4], OrdTy);
 828                 } else  {
 829                         order_succ = ConstantInt::get(OrdTy,
 830                                                         (int) AtomicOrderingCABI::seq_cst);
 831                         order_fail = ConstantInt::get(OrdTy,
 832                                                         (int) AtomicOrderingCABI::seq_cst);
 833                 }
 834
 835                 Value *args[] = {Addr, CmpOperand, NewOperand,
 836                                                         order_succ, order_fail, position};
 837
 838                 Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
 839                 ReplaceInstWithInst(CI, funcInst);
 840
 841                 return true;
 842         } else if ( funName.contains("compare_exchange_strong") ||
 843                                 funName.contains("compare_exchange_weak") ) {
 844                 Value *Addr = IRB.CreatePointerCast(OrigPtr, PtrTy);
 845                 Value *CmpOperand = IRB.CreatePointerCast(parameters[1], PtrTy);
 846                 Value *NewOperand = IRB.CreateBitOrPointerCast(parameters[2], Ty);
 847
 848                 Value *order_succ, *order_fail;
 849                 order_succ = IRB.CreateBitOrPointerCast(parameters[3], OrdTy);
 850                 order_fail = IRB.CreateBitOrPointerCast(parameters[4], OrdTy);
 851
 852                 Value *args[] = {Addr, CmpOperand, NewOperand,
 853                                                         order_succ, order_fail, position};
 854                 Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
 855                 ReplaceInstWithInst(CI, funcInst);
 856
 857                 return true;
 858         }
 859
 860         return false;
 861 }
 862
 863 int CDSPass::getMemoryAccessFuncIndex(Value *Addr,
 864                                                                                 const DataLayout &DL) {
 865         Type *OrigPtrTy = Addr->getType();
 866         Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
 867         assert(OrigTy->isSized());
 868         uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 869         if (TypeSize != 8  && TypeSize != 16 &&
 870                 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 871                 NumAccessesWithBadSize++;
 872                 // Ignore all unusual sizes.
 873                 return -1;
 874         }
 875         size_t Idx = countTrailingZeros(TypeSize / 8);
 876         //assert(Idx < kNumberOfAccessSizes);
 877         if (Idx >= kNumberOfAccessSizes) {
 878                 return -1;
 879         }
 880         return Idx;
 881 }
 882
 883
 884 char CDSPass::ID = 0;
 885
 886 // Automatically enable the pass.
 887 static void registerCDSPass(const PassManagerBuilder &,
 888                                                         legacy::PassManagerBase &PM) {
 889         PM.add(new CDSPass());
 890 }
 891
 892 /* Enable the pass when opt level is greater than 0 */
 893 static RegisterStandardPasses
 894         RegisterMyPass1(PassManagerBuilder::EP_OptimizerLast,
 895 registerCDSPass);
 896
 897 /* Enable the pass when opt level is 0 */
 898 static RegisterStandardPasses
 899         RegisterMyPass2(PassManagerBuilder::EP_EnabledOnOptLevel0,
 900 registerCDSPass);