Fix missed normal loads/stores
[c11llvm.git] / CDSPass.cpp
index ecbb9576548d0d1fed211fd6e6e924012358292e..a33738a96d26880395afc7beaf3b284e1bf1e705 100644 (file)
 #include "llvm/ADT/SmallString.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/Analysis/CaptureTracking.h"
+#include "llvm/Analysis/LoopInfo.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
-// #include "llvm/Transforms/Utils/ModuleUtils.h"
 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
 #include <vector>
 
 using namespace llvm;
 
+#define CDS_DEBUG
 #define DEBUG_TYPE "CDS"
 #include <llvm/IR/DebugLoc.h>
 
-Value *getPosition( Instruction * I, IRBuilder <> IRB, bool print = false)
+static inline Value *getPosition( Instruction * I, IRBuilder <> IRB, bool print = false)
 {
        const DebugLoc & debug_location = I->getDebugLoc ();
        std::string position_string;
@@ -66,6 +68,21 @@ Value *getPosition( Instruction * I, IRBuilder <> IRB, bool print = false)
        return IRB.CreateGlobalStringPtr (position_string);
 }
 
+static inline bool checkSignature(Function * func, Value * args[]) {
+       FunctionType * FType = func->getFunctionType();
+       for (unsigned i = 0 ; i < FType->getNumParams(); i++) {
+               if (FType->getParamType(i) != args[i]->getType()) {
+#ifdef CDS_DEBUG
+                       errs() << "expects: " << *FType->getParamType(i)
+                                       << "\tbut receives: " << *args[i]->getType() << "\n";
+#endif
+                       return false;
+               }
+       }
+
+       return true;
+}
+
 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
 STATISTIC(NumOmittedReadsBeforeWrite,
@@ -96,8 +113,8 @@ int getAtomicOrderIndex(AtomicOrdering order) {
        switch (order) {
                case AtomicOrdering::Monotonic: 
                        return (int)AtomicOrderingCABI::relaxed;
-               //  case AtomicOrdering::Consume:         // not specified yet
-               //    return AtomicOrderingCABI::consume;
+               //case AtomicOrdering::Consume:         // not specified yet
+               //      return AtomicOrderingCABI::consume;
                case AtomicOrdering::Acquire: 
                        return (int)AtomicOrderingCABI::acquire;
                case AtomicOrdering::Release: 
@@ -153,7 +170,7 @@ static Function * checkCDSPassInterfaceFunction(Constant *FuncOrBitcast) {
        if (isa<Function>(FuncOrBitcast))
                return cast<Function>(FuncOrBitcast);
        FuncOrBitcast->print(errs());
-       errs() << '\n';
+       errs() << "\n";
        std::string Err;
        raw_string_ostream Stream(Err);
        Stream << "CDSPass interface function redefined: " << *FuncOrBitcast;
@@ -172,14 +189,17 @@ namespace {
                void initializeCallbacks(Module &M);
                bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
                bool instrumentVolatile(Instruction *I, const DataLayout &DL);
+               bool instrumentMemIntrinsic(Instruction *I);
                bool isAtomicCall(Instruction *I);
                bool instrumentAtomic(Instruction *I, const DataLayout &DL);
                bool instrumentAtomicCall(CallInst *CI, const DataLayout &DL);
+               bool shouldInstrumentBeforeAtomics(Instruction *I);
                void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
                                                                                        SmallVectorImpl<Instruction *> &All,
                                                                                        const DataLayout &DL);
                bool addrPointsToConstantData(Value *Addr);
                int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
+               bool instrumentLoops(Function &F);
 
                Function * CDSFuncEntry;
                Function * CDSFuncExit;
@@ -195,7 +215,8 @@ namespace {
                Function * CDSAtomicCAS_V1[kNumberOfAccessSizes];
                Function * CDSAtomicCAS_V2[kNumberOfAccessSizes];
                Function * CDSAtomicThreadFence;
-               Function * CDSCtorFunction;
+               Function * MemmoveFn, * MemcpyFn, * MemsetFn;
+               // Function * CDSCtorFunction;
 
                std::vector<StringRef> AtomicFuncNames;
                std::vector<StringRef> PartialAtomicFuncNames;
@@ -208,8 +229,12 @@ StringRef CDSPass::getPassName() const {
 
 void CDSPass::initializeCallbacks(Module &M) {
        LLVMContext &Ctx = M.getContext();
+       AttributeList Attr;
+       Attr = Attr.addAttribute(Ctx, AttributeList::FunctionIndex,
+                       Attribute::NoUnwind);
 
        Type * Int1Ty = Type::getInt1Ty(Ctx);
+       Type * Int32Ty = Type::getInt32Ty(Ctx);
        OrdTy = Type::getInt32Ty(Ctx);
 
        Int8PtrTy  = Type::getInt8PtrTy(Ctx);
@@ -221,10 +246,10 @@ void CDSPass::initializeCallbacks(Module &M) {
 
        CDSFuncEntry = checkCDSPassInterfaceFunction(
                                                M.getOrInsertFunction("cds_func_entry", 
-                                               VoidTy, Int8PtrTy));
+                                               Attr, VoidTy, Int8PtrTy));
        CDSFuncExit = checkCDSPassInterfaceFunction(
                                                M.getOrInsertFunction("cds_func_exit", 
-                                               VoidTy, Int8PtrTy));
+                                               Attr, VoidTy, Int8PtrTy));
 
        // Get the function to call from our untime library.
        for (unsigned i = 0; i < kNumberOfAccessSizes; i++) {
@@ -248,24 +273,24 @@ void CDSPass::initializeCallbacks(Module &M) {
                SmallString<32> AtomicStoreName("cds_atomic_store" + BitSizeStr);
 
                CDSLoad[i]  = checkCDSPassInterfaceFunction(
-                                                       M.getOrInsertFunction(LoadName, VoidTy, PtrTy));
+                                                       M.getOrInsertFunction(LoadName, Attr, VoidTy, Int8PtrTy));
                CDSStore[i] = checkCDSPassInterfaceFunction(
-                                                       M.getOrInsertFunction(StoreName, VoidTy, PtrTy));
+                                                       M.getOrInsertFunction(StoreName, Attr, VoidTy, Int8PtrTy));
                CDSVolatileLoad[i]  = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(VolatileLoadName,
-                                                               Ty, PtrTy, Int8PtrTy));
+                                                               Attr, Ty, PtrTy, Int8PtrTy));
                CDSVolatileStore[i] = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(VolatileStoreName, 
-                                                               VoidTy, PtrTy, Ty, Int8PtrTy));
+                                                               Attr, VoidTy, PtrTy, Ty, Int8PtrTy));
                CDSAtomicInit[i] = checkCDSPassInterfaceFunction(
                                                        M.getOrInsertFunction(AtomicInitName, 
-                                                       VoidTy, PtrTy, Ty, Int8PtrTy));
+                                                       Attr, VoidTy, PtrTy, Ty, Int8PtrTy));
                CDSAtomicLoad[i]  = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(AtomicLoadName, 
-                                                               Ty, PtrTy, OrdTy, Int8PtrTy));
+                                                               Attr, Ty, PtrTy, OrdTy, Int8PtrTy));
                CDSAtomicStore[i] = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(AtomicStoreName, 
-                                                               VoidTy, PtrTy, Ty, OrdTy, Int8PtrTy));
+                                                               Attr, VoidTy, PtrTy, Ty, OrdTy, Int8PtrTy));
 
                for (int op = AtomicRMWInst::FIRST_BINOP; 
                        op <= AtomicRMWInst::LAST_BINOP; ++op) {
@@ -290,7 +315,7 @@ void CDSPass::initializeCallbacks(Module &M) {
                        SmallString<32> AtomicRMWName("cds_atomic" + NamePart + BitSizeStr);
                        CDSAtomicRMW[op][i] = checkCDSPassInterfaceFunction(
                                                                        M.getOrInsertFunction(AtomicRMWName, 
-                                                                       Ty, PtrTy, Ty, OrdTy, Int8PtrTy));
+                                                                       Attr, Ty, PtrTy, Ty, OrdTy, Int8PtrTy));
                }
 
                // only supportes strong version
@@ -298,15 +323,24 @@ void CDSPass::initializeCallbacks(Module &M) {
                SmallString<32> AtomicCASName_V2("cds_atomic_compare_exchange" + BitSizeStr + "_v2");
                CDSAtomicCAS_V1[i] = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(AtomicCASName_V1, 
-                                                               Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, Int8PtrTy));
+                                                               Attr, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, Int8PtrTy));
                CDSAtomicCAS_V2[i] = checkCDSPassInterfaceFunction(
                                                                M.getOrInsertFunction(AtomicCASName_V2, 
-                                                               Int1Ty, PtrTy, PtrTy, Ty, OrdTy, OrdTy, Int8PtrTy));
+                                                               Attr, Int1Ty, PtrTy, PtrTy, Ty, OrdTy, OrdTy, Int8PtrTy));
        }
 
        CDSAtomicThreadFence = checkCDSPassInterfaceFunction(
-                                                               M.getOrInsertFunction("cds_atomic_thread_fence", 
-                                                               VoidTy, OrdTy, Int8PtrTy));
+                       M.getOrInsertFunction("cds_atomic_thread_fence", Attr, VoidTy, OrdTy, Int8PtrTy));
+
+       MemmoveFn = checkCDSPassInterfaceFunction(
+                                       M.getOrInsertFunction("memmove", Attr, Int8PtrTy, Int8PtrTy,
+                                       Int8PtrTy, IntPtrTy));
+       MemcpyFn = checkCDSPassInterfaceFunction(
+                                       M.getOrInsertFunction("memcpy", Attr, Int8PtrTy, Int8PtrTy,
+                                       Int8PtrTy, IntPtrTy));
+       MemsetFn = checkCDSPassInterfaceFunction(
+                                       M.getOrInsertFunction("memset", Attr, Int8PtrTy, Int8PtrTy,
+                                       Int32Ty, IntPtrTy));
 }
 
 bool CDSPass::doInitialization(Module &M) {
@@ -396,6 +430,32 @@ bool CDSPass::addrPointsToConstantData(Value *Addr) {
        return false;
 }
 
+bool CDSPass::shouldInstrumentBeforeAtomics(Instruction * Inst) {
+       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
+               AtomicOrdering ordering = LI->getOrdering();
+               if ( isAtLeastOrStrongerThan(ordering, AtomicOrdering::Acquire) )
+                       return true;
+       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+               AtomicOrdering ordering = SI->getOrdering();
+               if ( isAtLeastOrStrongerThan(ordering, AtomicOrdering::Acquire) )
+                       return true;
+       } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(Inst)) {
+               AtomicOrdering ordering = RMWI->getOrdering();
+               if ( isAtLeastOrStrongerThan(ordering, AtomicOrdering::Acquire) )
+                       return true;
+       } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(Inst)) {
+               AtomicOrdering ordering = CASI->getSuccessOrdering();
+               if ( isAtLeastOrStrongerThan(ordering, AtomicOrdering::Acquire) )
+                       return true;
+       } else if (FenceInst *FI = dyn_cast<FenceInst>(Inst)) {
+               AtomicOrdering ordering = FI->getOrdering();
+               if ( isAtLeastOrStrongerThan(ordering, AtomicOrdering::Acquire) )
+                       return true;
+       }
+
+       return false;
+}
+
 void CDSPass::chooseInstructionsToInstrument(
        SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
        const DataLayout &DL) {
@@ -449,28 +509,35 @@ void CDSPass::InsertRuntimeIgnores(Function &F) {
 }*/
 
 bool CDSPass::runOnFunction(Function &F) {
-       if (F.getName() == "main") {
-               F.setName("user_main");
-               errs() << "main replaced by user_main\n";
-       }
-
        initializeCallbacks( *F.getParent() );
        SmallVector<Instruction*, 8> AllLoadsAndStores;
        SmallVector<Instruction*, 8> LocalLoadsAndStores;
        SmallVector<Instruction*, 8> VolatileLoadsAndStores;
        SmallVector<Instruction*, 8> AtomicAccesses;
+       SmallVector<Instruction*, 8> MemIntrinCalls;
 
        bool Res = false;
-       bool HasCall = false;
        bool HasAtomic = false;
        bool HasVolatile = false;
        const DataLayout &DL = F.getParent()->getDataLayout();
 
+       // instrumentLoops(F);
+
        for (auto &BB : F) {
                for (auto &Inst : BB) {
-                       if ( (&Inst)->isAtomic() || isAtomicCall(&Inst) ) {
+                       if ( (&Inst)->isAtomic() ) {
+                               AtomicAccesses.push_back(&Inst);
+                               HasAtomic = true;
+
+                               if (shouldInstrumentBeforeAtomics(&Inst)) {
+                                       chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
+                                               DL);
+                               }
+                       } else if (isAtomicCall(&Inst) ) {
                                AtomicAccesses.push_back(&Inst);
                                HasAtomic = true;
+                               chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
+                                       DL);
                        } else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst)) {
                                LoadInst *LI = dyn_cast<LoadInst>(&Inst);
                                StoreInst *SI = dyn_cast<StoreInst>(&Inst);
@@ -482,15 +549,15 @@ bool CDSPass::runOnFunction(Function &F) {
                                } else
                                        LocalLoadsAndStores.push_back(&Inst);
                        } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
-                               /* TODO: To be added
-                               if (CallInst *CI = dyn_cast<CallInst>(&Inst))
-                                       maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
                                if (isa<MemIntrinsic>(Inst))
                                        MemIntrinCalls.push_back(&Inst);
-                               HasCalls = true;
+
+                               /*if (CallInst *CI = dyn_cast<CallInst>(&Inst))
+                                       maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
+                               */
+
                                chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
                                        DL);
-                               */
                        }
                }
 
@@ -509,13 +576,11 @@ bool CDSPass::runOnFunction(Function &F) {
                Res |= instrumentAtomic(Inst, DL);
        }
 
-       /* TODO
        for (auto Inst : MemIntrinCalls) {
                Res |= instrumentMemIntrinsic(Inst);
        }
-       */
 
-       // Only instrument functions that contain atomics or volatiles
+       // Instrument function entry and exit for functions containing atomics or volatiles
        if (Res && ( HasAtomic || HasVolatile) ) {
                IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
                /* Unused for now
@@ -550,52 +615,46 @@ bool CDSPass::instrumentLoadOrStore(Instruction *I,
        // As such they cannot have regular uses like an instrumentation function and
        // it makes no sense to track them as memory.
        if (Addr->isSwiftError())
-       return false;
+               return false;
 
        int Idx = getMemoryAccessFuncIndex(Addr, DL);
        if (Idx < 0)
                return false;
 
-//  not supported by CDS yet
-/*  if (IsWrite && isVtableAccess(I)) {
-    LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
-    Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
-    // StoredValue may be a vector type if we are storing several vptrs at once.
-    // In this case, just take the first element of the vector since this is
-    // enough to find vptr races.
-    if (isa<VectorType>(StoredValue->getType()))
-      StoredValue = IRB.CreateExtractElement(
-          StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
-    if (StoredValue->getType()->isIntegerTy())
-      StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
-    // Call TsanVptrUpdate.
-    IRB.CreateCall(TsanVptrUpdate,
-                   {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
-                    IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
-    NumInstrumentedVtableWrites++;
-    return true;
-  }
-
-  if (!IsWrite && isVtableAccess(I)) {
-    IRB.CreateCall(TsanVptrLoad,
-                   IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
-    NumInstrumentedVtableReads++;
-    return true;
-  }
-*/
-
-       Value *OnAccessFunc = nullptr;
-       OnAccessFunc = IsWrite ? CDSStore[Idx] : CDSLoad[Idx];
-
-       Type *ArgType = IRB.CreatePointerCast(Addr, Addr->getType())->getType();
+       if (IsWrite && isVtableAccess(I)) {
+               /* TODO
+               LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
+               Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
+               // StoredValue may be a vector type if we are storing several vptrs at once.
+               // In this case, just take the first element of the vector since this is
+               // enough to find vptr races.
+               if (isa<VectorType>(StoredValue->getType()))
+                       StoredValue = IRB.CreateExtractElement(
+                                       StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
+               if (StoredValue->getType()->isIntegerTy())
+                       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
+               // Call TsanVptrUpdate.
+               IRB.CreateCall(TsanVptrUpdate,
+                                               {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
+                                                       IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
+               NumInstrumentedVtableWrites++;
+               */
+               return true;
+       }
 
-       if ( ArgType != Int8PtrTy && ArgType != Int16PtrTy && 
-                       ArgType != Int32PtrTy && ArgType != Int64PtrTy ) {
-               // if other types of load or stores are passed in
-               return false;   
+       if (!IsWrite && isVtableAccess(I)) {
+               /* TODO
+               IRB.CreateCall(TsanVptrLoad,
+                                                IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
+               NumInstrumentedVtableReads++;
+               */
+               return true;
        }
 
-       IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, Addr->getType()));
+       // TODO: unaligned reads and writes
+       Value *OnAccessFunc = nullptr;
+       OnAccessFunc = IsWrite ? CDSStore[Idx] : CDSLoad[Idx];
+       IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
        if (IsWrite) NumInstrumentedWrites++;
        else         NumInstrumentedReads++;
        return true;
@@ -606,26 +665,34 @@ bool CDSPass::instrumentVolatile(Instruction * I, const DataLayout &DL) {
        Value *position = getPosition(I, IRB);
 
        if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
-               assert( LI->isVolatile() );
                Value *Addr = LI->getPointerOperand();
                int Idx=getMemoryAccessFuncIndex(Addr, DL);
                if (Idx < 0)
                        return false;
+               const unsigned ByteSize = 1U << Idx;
+               const unsigned BitSize = ByteSize * 8;
+               Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
+               Type *PtrTy = Ty->getPointerTo();
+               Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), position};
 
-               Value *args[] = {Addr, position};
-               Instruction* funcInst = CallInst::Create(CDSVolatileLoad[Idx], args);
-               ReplaceInstWithInst(LI, funcInst);
+               Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
+               Value *C = IRB.CreateCall(CDSVolatileLoad[Idx], Args);
+               Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
+               I->replaceAllUsesWith(Cast);
        } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
-               assert( SI->isVolatile() );
                Value *Addr = SI->getPointerOperand();
                int Idx=getMemoryAccessFuncIndex(Addr, DL);
                if (Idx < 0)
                        return false;
-
-               Value *val = SI->getValueOperand();
-               Value *args[] = {Addr, val, position};
-               Instruction* funcInst = CallInst::Create(CDSVolatileStore[Idx], args);
-               ReplaceInstWithInst(SI, funcInst);
+               const unsigned ByteSize = 1U << Idx;
+               const unsigned BitSize = ByteSize * 8;
+               Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
+               Type *PtrTy = Ty->getPointerTo();
+               Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
+                                         IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
+                                         position};
+               CallInst *C = CallInst::Create(CDSVolatileStore[Idx], Args);
+               ReplaceInstWithInst(I, C);
        } else {
                return false;
        }
@@ -633,6 +700,26 @@ bool CDSPass::instrumentVolatile(Instruction * I, const DataLayout &DL) {
        return true;
 }
 
+bool CDSPass::instrumentMemIntrinsic(Instruction *I) {
+       IRBuilder<> IRB(I);
+       if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
+               IRB.CreateCall(
+                       MemsetFn,
+                       {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
+                        IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
+                        IRB.CreateIntCast(M->getArgOperand(2), IntPtrTy, false)});
+               I->eraseFromParent();
+       } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
+               IRB.CreateCall(
+                       isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
+                       {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
+                        IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
+                        IRB.CreateIntCast(M->getArgOperand(2), IntPtrTy, false)});
+               I->eraseFromParent();
+       }
+       return false;
+}
+
 bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
        IRBuilder<> IRB(I);
 
@@ -641,7 +728,6 @@ bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
        }
 
        Value *position = getPosition(I, IRB);
-
        if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
                Value *Addr = LI->getPointerOperand();
                int Idx=getMemoryAccessFuncIndex(Addr, DL);
@@ -650,8 +736,8 @@ bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
 
                int atomic_order_index = getAtomicOrderIndex(LI->getOrdering());
                Value *order = ConstantInt::get(OrdTy, atomic_order_index);
-               Value *args[] = {Addr, order, position};
-               Instruction* funcInst = CallInst::Create(CDSAtomicLoad[Idx], args);
+               Value *Args[] = {Addr, order, position};
+               Instruction* funcInst = CallInst::Create(CDSAtomicLoad[Idx], Args);
                ReplaceInstWithInst(LI, funcInst);
        } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
                Value *Addr = SI->getPointerOperand();
@@ -662,8 +748,8 @@ bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
                int atomic_order_index = getAtomicOrderIndex(SI->getOrdering());
                Value *val = SI->getValueOperand();
                Value *order = ConstantInt::get(OrdTy, atomic_order_index);
-               Value *args[] = {Addr, val, order, position};
-               Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
+               Value *Args[] = {Addr, val, order, position};
+               Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], Args);
                ReplaceInstWithInst(SI, funcInst);
        } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
                Value *Addr = RMWI->getPointerOperand();
@@ -674,8 +760,8 @@ bool CDSPass::instrumentAtomic(Instruction * I, const DataLayout &DL) {
                int atomic_order_index = getAtomicOrderIndex(RMWI->getOrdering());
                Value *val = RMWI->getValOperand();
                Value *order = ConstantInt::get(OrdTy, atomic_order_index);
-               Value *args[] = {Addr, val, order, position};
-               Instruction* funcInst = CallInst::Create(CDSAtomicRMW[RMWI->getOperation()][Idx], args);
+               Value *Args[] = {Addr, val, order, position};
+               Instruction* funcInst = CallInst::Create(CDSAtomicRMW[RMWI->getOperation()][Idx], Args);
                ReplaceInstWithInst(RMWI, funcInst);
        } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
                IRBuilder<> IRB(CASI);
@@ -738,7 +824,7 @@ bool CDSPass::isAtomicCall(Instruction *I) {
 
                StringRef funName = fun->getName();
 
-               // todo: come up with better rules for function name checking
+               // TODO: come up with better rules for function name checking
                for (StringRef name : AtomicFuncNames) {
                        if ( funName.contains(name) ) 
                                return true;
@@ -795,9 +881,11 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
 
                Value *args[] = {ptr, val, position};
 
+               if (!checkSignature(CDSAtomicInit[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicInit[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
-
                return true;
        }
 
@@ -813,7 +901,10 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
                        order = ConstantInt::get(OrdTy, 
                                                        (int) AtomicOrderingCABI::seq_cst);
                Value *args[] = {ptr, order, position};
-               
+
+               if (!checkSignature(CDSAtomicLoad[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicLoad[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
@@ -825,10 +916,14 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
                Value *order = IRB.CreateBitOrPointerCast(parameters[1], OrdTy);
                Value *args[] = {ptr, order, position};
 
+               // Without this check, gdax does not compile :(
                if (!CI->getType()->isPointerTy()) {
                        return false;   
                } 
 
+               if (!checkSignature(CDSAtomicLoad[Idx], args))
+                       return false;
+
                CallInst *funcInst = IRB.CreateCall(CDSAtomicLoad[Idx], args);
                Value *RetVal = IRB.CreateIntToPtr(funcInst, CI->getType());
 
@@ -852,17 +947,23 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
                        order = ConstantInt::get(OrdTy, 
                                                        (int) AtomicOrderingCABI::seq_cst);
                Value *args[] = {ptr, val, order, position};
-               
+
+               if (!checkSignature(CDSAtomicStore[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
                return true;
        } else if (funName.contains("atomic") && 
                                        funName.contains("store") ) {
-               // does this version of call always have an atomic order as an argument?
-               Value *OrigVal = parameters[1];
+               // Does this version of call always have an atomic order as an argument?
+               if (parameters.size() < 3)
+                       return false;
 
+               Value *OrigVal = parameters[1];
                Value *ptr = IRB.CreatePointerCast(OrigPtr, PtrTy);
+
                Value *val;
                if (OrigVal->getType()->isPtrOrPtrVectorTy())
                        val = IRB.CreatePointerCast(OrigVal, Ty);
@@ -872,6 +973,9 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
                Value *order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
                Value *args[] = {ptr, val, order, position};
 
+               if (!checkSignature(CDSAtomicStore[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicStore[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
@@ -882,10 +986,6 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
        if (funName.contains("atomic_fetch_") || 
                funName.contains("atomic_exchange")) {
 
-               /* TODO: implement stricter function name checking */
-               if (funName.contains("non"))
-                       return false;
-
                bool isExplicit = funName.contains("_explicit");
                Value *OrigVal = parameters[1];
 
@@ -921,7 +1021,10 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
                        order = ConstantInt::get(OrdTy, 
                                                        (int) AtomicOrderingCABI::seq_cst);
                Value *args[] = {ptr, val, order, position};
-               
+
+               if (!checkSignature(CDSAtomicRMW[op][Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicRMW[op][Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
@@ -934,7 +1037,13 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
        } else if (funName.contains("exchange") &&
                        !funName.contains("compare_exchange") ) {
                if (CI->getType()->isPointerTy()) {
-                       // Can not deal with this now
+                       /**
+                        * TODO: instrument the following case
+                        * mcs-lock.h
+                        * std::atomic<struct T *> m_tail;
+                        * struct T * me;
+                        * struct T * pred = m_tail.exchange(me, memory_order_*);
+                        */
                        errs() << "atomic exchange captured. Not implemented yet. ";
                        errs() << "See source file :";
                        getPosition(CI, IRB, true);
@@ -953,10 +1062,16 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
 
                Value *order = IRB.CreateBitOrPointerCast(parameters[2], OrdTy);
                Value *args[] = {ptr, val, order, position};
+
                int op = AtomicRMWInst::Xchg;
-               
+
+               if (!checkSignature(CDSAtomicRMW[op][Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicRMW[op][Idx], args);
                ReplaceInstWithInst(CI, funcInst);
+
+               return true;
        }
 
        /* atomic_compare_exchange_*; 
@@ -993,7 +1108,10 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
 
                Value *args[] = {Addr, CmpOperand, NewOperand, 
                                                        order_succ, order_fail, position};
-               
+
+               if (!checkSignature(CDSAtomicCAS_V2[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
@@ -1021,6 +1139,10 @@ bool CDSPass::instrumentAtomicCall(CallInst *CI, const DataLayout &DL) {
 
                Value *args[] = {Addr, CmpOperand, NewOperand, 
                                                        order_succ, order_fail, position};
+
+               if (!checkSignature(CDSAtomicCAS_V2[Idx], args))
+                       return false;
+
                Instruction* funcInst = CallInst::Create(CDSAtomicCAS_V2[Idx], args);
                ReplaceInstWithInst(CI, funcInst);
 
@@ -1050,6 +1172,58 @@ int CDSPass::getMemoryAccessFuncIndex(Value *Addr,
        return Idx;
 }
 
+bool CDSPass::instrumentLoops(Function &F)
+{
+       DominatorTree DT(F);
+       LoopInfo LI(DT);
+
+       SmallVector<Loop *, 4> Loops = LI.getLoopsInPreorder();
+       bool instrumented = false;
+
+       // Do a post-order traversal of the loops so that counter updates can be
+       // iteratively hoisted outside the loop nest.
+       for (auto *Loop : llvm::reverse(Loops)) {
+               bool instrument_loop = false;
+
+               // Iterator over loop blocks and search for atomics and volatiles
+               Loop::block_iterator it;
+               for (it = Loop->block_begin(); it != Loop->block_end(); it++) {
+                       BasicBlock * block = *it;
+                       for (auto &Inst : *block) {
+                               if ( (&Inst)->isAtomic() ) {
+                                       instrument_loop = true;
+                                       break;
+                               } else if (isAtomicCall(&Inst)) {
+                                       instrument_loop = true;
+                                       break;
+                               } else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst)) {
+                                       LoadInst *LI = dyn_cast<LoadInst>(&Inst);
+                                       StoreInst *SI = dyn_cast<StoreInst>(&Inst);
+                                       bool isVolatile = ( LI ? LI->isVolatile() : SI->isVolatile() );
+
+                                       if (isVolatile) {
+                                               instrument_loop = true;
+                                               break;
+                                       }
+                               }
+                       }
+
+                       if (instrument_loop)
+                               break;
+               }
+
+               if (instrument_loop) {
+                       // TODO: what to instrument?
+                       errs() << "Function: " << F.getName() << "\n";
+                       BasicBlock * header = Loop->getHeader();
+                       header->dump();
+
+                       instrumented = true;
+               }
+       }
+
+       return instrumented;
+}
 
 char CDSPass::ID = 0;