From: Peizhao Ou <peizhaoo@uci.edu>
Date: Thu, 5 Apr 2018 22:42:56 +0000 (-0700)
Subject: Reverts unnecessary changes
X-Git-Url: http://plrg.eecs.uci.edu/git/?a=commitdiff_plain;h=refs%2Fheads%2Frelaxed-stores-insert-dmb;p=oota-llvm.git

Reverts unnecessary changes
---

diff --git a/include/llvm/CodeGen/MachineBasicBlock.h b/include/llvm/CodeGen/MachineBasicBlock.h
index fd786f68fb6..3d58c499823 100644
--- a/include/llvm/CodeGen/MachineBasicBlock.h
+++ b/include/llvm/CodeGen/MachineBasicBlock.h
@@ -124,10 +124,6 @@ private:
   /// is only computed once and is cached.
   mutable MCSymbol *CachedMCSymbol = nullptr;
 
-  // XXX-update: A flag that checks whether we can eliminate this machine basic
-  // block.
-  bool canEliminateMachineBB;
-
   // Intrusive list support
   MachineBasicBlock() {}
 
@@ -139,15 +135,6 @@ private:
   friend class MachineFunction;
 
 public:
-  // XXX-update:
-  void disableCanEliminateMachineBB() {
-    canEliminateMachineBB = false;
-  }
-
-  bool getCanEliminateMachineBB() {
-    return canEliminateMachineBB;
-  }
-
   /// Return the LLVM basic block that this instance corresponded to originally.
   /// Note that this may be NULL if this instance does not correspond directly
   /// to an LLVM basic block.
diff --git a/include/llvm/IR/BasicBlock.h b/include/llvm/IR/BasicBlock.h
index 94edb547f53..c6b54d308ce 100644
--- a/include/llvm/IR/BasicBlock.h
+++ b/include/llvm/IR/BasicBlock.h
@@ -59,9 +59,6 @@ private:
   InstListType InstList;
   Function *Parent;
 
-  // XXX-update: A flag that checks whether we can eliminate this block.
-  bool canEliminateBlock;
-
   void setParent(Function *parent);
   friend class SymbolTableListTraits<BasicBlock>;
 
@@ -77,16 +74,6 @@ private:
                       Function *Parent = nullptr,
                       BasicBlock *InsertBefore = nullptr);
 public:
-  // XXX-update:
-  void disableCanEliminateBlock() {
-    canEliminateBlock = false;
-  }
-
-  bool getCanEliminateBlock() {
-    return canEliminateBlock;
-  }
-
-
   /// \brief Get the context in which this basic block lives.
   LLVMContext &getContext() const;
 
diff --git a/include/llvm/IR/Instructions.h b/include/llvm/IR/Instructions.h
index cc6c25974cd..28e1fd90fdf 100644
--- a/include/llvm/IR/Instructions.h
+++ b/include/llvm/IR/Instructions.h
@@ -228,14 +228,6 @@ public:
   LoadInst(Value *Ptr, const char *NameStr, bool isVolatile,
            BasicBlock *InsertAtEnd);
 
-  bool getHasSubsequentAcqlRMW() {
-    return hasSubsequentAcqlRMW_;
-  }
-
-  void setHasSubsequentAcqlRMW(bool val) {
-    hasSubsequentAcqlRMW_ = val;
-  }
-
   /// isVolatile - Return true if this is a load from a volatile memory
   /// location.
   ///
@@ -314,8 +306,6 @@ private:
   void setInstructionSubclassData(unsigned short D) {
     Instruction::setInstructionSubclassData(D);
   }
-
-  bool hasSubsequentAcqlRMW_;
 };
 
 //===----------------------------------------------------------------------===//
diff --git a/lib/CodeGen/AtomicExpandPass.cpp b/lib/CodeGen/AtomicExpandPass.cpp
index 9647fb9e5d5..d12fdb24698 100644
--- a/lib/CodeGen/AtomicExpandPass.cpp
+++ b/lib/CodeGen/AtomicExpandPass.cpp
@@ -15,11 +15,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#include "llvm/ADT/SetOperations.h"
-#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/Analysis/MemoryLocation.h"
 #include "llvm/CodeGen/AtomicExpandUtils.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/Function.h"
@@ -28,13 +23,11 @@
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/Module.h"
-#include "llvm/IR/NoFolder.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Target/TargetLowering.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetSubtargetInfo.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
 
 using namespace llvm;
 
@@ -71,33 +64,6 @@ namespace {
     bool isIdempotentRMW(AtomicRMWInst *AI);
     bool simplifyIdempotentRMW(AtomicRMWInst *AI);
   };
-
-
-  // If 'LI' is a relaxed load, and it is immediately followed by a
-// atomic read-modify-write that has acq_rel parameter, we don't have to do
-// anything since the rmw serves as a natural barrier.
-void MarkRelaxedLoadBeforeAcqrelRMW(LoadInst* LI) {
-  auto* BB = LI->getParent();
-  auto BBI = LI->getIterator();
-  for (BBI++; BBI != BB->end(); BBI++) {
-    Instruction* CurInst = &*BBI;
-    if (!CurInst) {
-      return;
-    }
-    if (!CurInst->isAtomic()) {
-      continue;
-    }
-    auto* RMW = dyn_cast<AtomicRMWInst>(CurInst);
-    if (!RMW) {
-      return;
-    }
-    if (RMW->getOrdering() == AcquireRelease ||
-        RMW->getOrdering() == SequentiallyConsistent) {
-      LI->setHasSubsequentAcqlRMW(true);
-    }
-  }
-}
-
 }
 
 char AtomicExpand::ID = 0;
@@ -116,64 +82,15 @@ bool AtomicExpand::runOnFunction(Function &F) {
   TLI = TM->getSubtargetImpl(F)->getTargetLowering();
 
   SmallVector<Instruction *, 1> AtomicInsts;
-  SmallVector<LoadInst*, 1> MonotonicLoadInsts;
 
-  bool MadeChange = false;
   // Changing control-flow while iterating through it is a bad idea, so gather a
   // list of all atomic instructions before we start.
   for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
-    // XXX-update: For relaxed loads, change them to acquire. This includes
-    // relaxed loads, relaxed atomic RMW & relaxed atomic compare exchange.
-    if (I->isAtomic()) {
-      switch (I->getOpcode()) {
-        case Instruction::AtomicCmpXchg: {
-          // XXX-comment: AtomicCmpXchg in AArch64 will be translated to a
-          // conditional branch that contains the value of the load anyway, so
-          // we don't need to do anything.
-          /*
-          auto* CmpXchg = dyn_cast<AtomicCmpXchgInst>(&*I);
-          auto SuccOrdering = CmpXchg->getSuccessOrdering();
-          if (SuccOrdering == Monotonic) {
-            CmpXchg->setSuccessOrdering(Acquire);
-          } else if (SuccOrdering == Release) {
-            CmpXchg->setSuccessOrdering(AcquireRelease);
-          }
-          */
-          break;
-        }
-        case Instruction::AtomicRMW: {
-          // XXX-comment: Similar to AtomicCmpXchg. These instructions in
-          // AArch64 will be translated to a loop whose condition depends on the
-          // store status, which further depends on the load value.
-          /*
-          auto* RMW = dyn_cast<AtomicRMWInst>(&*I);
-          if (RMW->getOrdering() == Monotonic) {
-            RMW->setOrdering(Acquire);
-          }
-          */
-          break;
-        }
-        case Instruction::Load: {
-          auto* LI = dyn_cast<LoadInst>(&*I);
-          if (LI->getOrdering() == Monotonic) {
-            /*
-            DEBUG(dbgs() << "Transforming relaxed loads to acquire loads: "
-                         << *LI << '\n');
-            LI->setOrdering(Acquire);
-            */
-//            MonotonicLoadInsts.push_back(LI);
-            MarkRelaxedLoadBeforeAcqrelRMW(LI);
-          }
-          break;
-        }
-        default: {
-          break;
-        }
-      }
+    if (I->isAtomic())
       AtomicInsts.push_back(&*I);
-    }
   }
 
+  bool MadeChange = false;
   for (auto I : AtomicInsts) {
     auto LI = dyn_cast<LoadInst>(I);
     auto SI = dyn_cast<StoreInst>(I);
@@ -187,7 +104,7 @@ bool AtomicExpand::runOnFunction(Function &F) {
     if (TLI->getInsertFencesForAtomic()) {
       if (LI && isAtLeastAcquire(LI->getOrdering())) {
         FenceOrdering = LI->getOrdering();
-//        AddFakeConditionalBranch(
+        LI->setOrdering(Monotonic);
         IsStore = false;
         IsLoad = true;
       } else if (SI && isAtLeastRelease(SI->getOrdering())) {
@@ -255,7 +172,6 @@ bool AtomicExpand::runOnFunction(Function &F) {
       MadeChange |= expandAtomicCmpXchg(CASI);
     }
   }
-
   return MadeChange;
 }
 
diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp
index a56fec468cc..df5cac5a9f7 100644
--- a/lib/CodeGen/BranchFolding.cpp
+++ b/lib/CodeGen/BranchFolding.cpp
@@ -1116,12 +1116,6 @@ bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
   for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
        I != E; ) {
     MachineBasicBlock *MBB = &*I++;
-    // XXX-disabled: Don't optimize blocks that contain intentionally added fake
-    // conditional branch.
-    if (!MBB->getCanEliminateMachineBB()) {
-      continue;
-    }
-
     MadeChange |= OptimizeBlock(MBB);
 
     // If it is dead, remove it.
diff --git a/lib/CodeGen/CodeGenPrepare.cpp b/lib/CodeGen/CodeGenPrepare.cpp
index bb183b4d6dc..c8007a524e7 100644
--- a/lib/CodeGen/CodeGenPrepare.cpp
+++ b/lib/CodeGen/CodeGenPrepare.cpp
@@ -15,13 +15,9 @@
 
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/InstructionSimplify.h"
-#include "llvm/Analysis/MemoryLocation.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
@@ -34,12 +30,9 @@
 #include "llvm/IR/GetElementPtrTypeIterator.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InlineAsm.h"
-#include "llvm/IR/InstIterator.h"
-#include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/MDBuilder.h"
-#include "llvm/IR/NoFolder.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/IR/Statepoint.h"
 #include "llvm/IR/ValueHandle.h"
@@ -154,15 +147,10 @@ class TypePromotionTransaction;
     /// DataLayout for the Function being processed.
     const DataLayout *DL;
 
-    // XXX-comment:We need DominatorTree to figure out which instruction to
-    // taint.
-    DominatorTree *DT;
-
   public:
     static char ID; // Pass identification, replacement for typeid
     explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
-        : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr),
-        DT(nullptr) {
+        : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
         initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
       }
     bool runOnFunction(Function &F) override;
@@ -173,7 +161,6 @@ class TypePromotionTransaction;
       AU.addPreserved<DominatorTreeWrapperPass>();
       AU.addRequired<TargetLibraryInfoWrapperPass>();
       AU.addRequired<TargetTransformInfoWrapperPass>();
-      AU.addRequired<DominatorTreeWrapperPass>();
     }
 
   private:
@@ -208,1109 +195,20 @@ class TypePromotionTransaction;
 }
 
 char CodeGenPrepare::ID = 0;
-INITIALIZE_TM_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
-                   "Optimize for code generation", false, false)
-INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
-INITIALIZE_TM_PASS_END(CodeGenPrepare, "codegenprepare",
+INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
                    "Optimize for code generation", false, false)
 
 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
   return new CodeGenPrepare(TM);
 }
 
-namespace {
-
-bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal);
-Value* GetUntaintedAddress(Value* CurrentAddress);
-
-// The depth we trace down a variable to look for its dependence set.
-const unsigned kDependenceDepth = 4;
-
-// Recursively looks for variables that 'Val' depends on at the given depth
-// 'Depth', and adds them in 'DepSet'. If 'InsertOnlyLeafNodes' is true, only
-// inserts the leaf node values; otherwise, all visited nodes are included in
-// 'DepSet'. Note that constants will be ignored.
-template <typename SetType>
-void recursivelyFindDependence(SetType* DepSet, Value* Val,
-                               bool InsertOnlyLeafNodes = false,
-                               unsigned Depth = kDependenceDepth) {
-  if (Val == nullptr) {
-    return;
-  }
-  if (!InsertOnlyLeafNodes && !isa<Constant>(Val)) {
-    DepSet->insert(Val);
-  }
-  if (Depth == 0) {
-    // Cannot go deeper. Insert the leaf nodes.
-    if (InsertOnlyLeafNodes && !isa<Constant>(Val)) {
-      DepSet->insert(Val);
-    }
-    return;
-  }
-
-  // Go one step further to explore the dependence of the operands.
-  Instruction* I = nullptr;
-  if ((I = dyn_cast<Instruction>(Val))) {
-    if (isa<LoadInst>(I)) {
-      // A load is considerd the leaf load of the dependence tree. Done.
-      DepSet->insert(Val);
-      return;
-    } else if (I->isBinaryOp()) {
-      BinaryOperator* I = dyn_cast<BinaryOperator>(Val);
-      Value *Op0 = I->getOperand(0), *Op1 = I->getOperand(1);
-      recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
-      recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
-    } else if (I->isCast()) {
-      Value* Op0 = I->getOperand(0);
-      recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
-    } else if (I->getOpcode() == Instruction::Select) {
-      Value* Op0 = I->getOperand(0);
-      Value* Op1 = I->getOperand(1);
-      Value* Op2 = I->getOperand(2);
-      recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
-      recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
-      recursivelyFindDependence(DepSet, Op2, InsertOnlyLeafNodes, Depth - 1);
-    } else if (I->getOpcode() == Instruction::GetElementPtr) {
-      for (unsigned i = 0; i < I->getNumOperands(); i++) {
-        recursivelyFindDependence(DepSet, I->getOperand(i), InsertOnlyLeafNodes,
-                                  Depth - 1);
-      }
-    } else if (I->getOpcode() == Instruction::Store) {
-      auto* SI = dyn_cast<StoreInst>(Val);
-      recursivelyFindDependence(DepSet, SI->getPointerOperand(),
-                                InsertOnlyLeafNodes, Depth - 1);
-      recursivelyFindDependence(DepSet, SI->getValueOperand(),
-                                InsertOnlyLeafNodes, Depth - 1);
-    } else {
-      Value* Op0 = nullptr;
-      Value* Op1 = nullptr;
-      switch (I->getOpcode()) {
-        case Instruction::ICmp:
-        case Instruction::FCmp: {
-          Op0 = I->getOperand(0);
-          Op1 = I->getOperand(1);
-          recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes,
-                                    Depth - 1);
-          recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes,
-                                    Depth - 1);
-          break;
-        }
-        case Instruction::PHI: {
-          for (int i = 0; i < I->getNumOperands(); i++) {
-            auto* op = I->getOperand(i);
-            if (DepSet->count(op) == 0) {
-              recursivelyFindDependence(DepSet, I->getOperand(i),
-                                        InsertOnlyLeafNodes, Depth - 1);
-            }
-          }
-          break;
-        }
-        default: {
-          // Be conservative. Add it and be done with it.
-          DepSet->insert(Val);
-          return;
-        }
-      }
-    }
-  } else if (isa<Constant>(Val)) {
-    // Not interested in constant values. Done.
-    return;
-  } else {
-    // Be conservative. Add it and be done with it.
-    DepSet->insert(Val);
-    return;
-  }
-}
-
-// Helper function to create a Cast instruction.
-Value* createCast(IRBuilder<true, NoFolder>& Builder, Value* DepVal,
-                  Type* TargetIntegerType) {
-  Instruction::CastOps CastOp = Instruction::BitCast;
-  switch (DepVal->getType()->getTypeID()) {
-    case Type::IntegerTyID: {
-      assert(TargetIntegerType->getTypeID() == Type::IntegerTyID);
-      auto* FromType = dyn_cast<IntegerType>(DepVal->getType());
-      auto* ToType = dyn_cast<IntegerType>(TargetIntegerType);
-      assert(FromType && ToType);
-      if (FromType->getBitWidth() <= ToType->getBitWidth()) {
-        CastOp = Instruction::ZExt;
-      } else {
-        CastOp = Instruction::Trunc;
-      }
-      break;
-    }
-    case Type::FloatTyID:
-    case Type::DoubleTyID: {
-      CastOp = Instruction::FPToSI;
-      break;
-    }
-    case Type::PointerTyID: {
-      CastOp = Instruction::PtrToInt;
-      break;
-    }
-    default: { break; }
-  }
-
-  return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
-}
-
-// Given a value, if it's a tainted address, this function returns the
-// instruction that ORs the "dependence value" with the "original address".
-// Otherwise, returns nullptr.  This instruction is the first OR instruction
-// where one of its operand is an AND instruction with an operand being 0.
-//
-// E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
-// %0 = load i32, i32* @y, align 4, !tbaa !1
-// %cmp = icmp ne i32 %0, 42  // <== this is like the condition
-// %1 = sext i1 %cmp to i32
-// %2 = ptrtoint i32* @x to i32
-// %3 = and i32 %1, 0
-// %4 = or i32 %3, %2
-// %5 = inttoptr i32 %4 to i32*
-// store i32 1, i32* %5, align 4
-Instruction* getOrAddress(Value* CurrentAddress) {
-  // Is it a cast from integer to pointer type.
-  Instruction* OrAddress = nullptr;
-  Instruction* AndDep = nullptr;
-  Instruction* CastToInt = nullptr;
-  Value* ActualAddress = nullptr;
-  Constant* ZeroConst = nullptr;
-
-  const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
-  if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
-    // Is it an OR instruction: %1 = or %and, %actualAddress.
-    if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
-        OrAddress->getOpcode() == Instruction::Or) {
-      // The first operand should be and AND instruction.
-      AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
-      if (AndDep && AndDep->getOpcode() == Instruction::And) {
-        // Also make sure its first operand of the "AND" is 0, or the "AND" is
-        // marked explicitly by "NoInstCombine".
-        if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
-            ZeroConst->isNullValue()) {
-          return OrAddress;
-        }
-      }
-    }
-  }
-  // Looks like it's not been tainted.
-  return nullptr;
-}
-
-// Given a value, if it's a tainted address, this function returns the
-// instruction that taints the "dependence value". Otherwise, returns nullptr.
-// This instruction is the last AND instruction where one of its operand is 0.
-// E.g., it returns '%3' given 'CurrentAddress' is '%5'.
-// %0 = load i32, i32* @y, align 4, !tbaa !1
-// %cmp = icmp ne i32 %0, 42  // <== this is like the condition
-// %1 = sext i1 %cmp to i32
-// %2 = ptrtoint i32* @x to i32
-// %3 = and i32 %1, 0
-// %4 = or i32 %3, %2
-// %5 = inttoptr i32 %4 to i32*
-// store i32 1, i32* %5, align 4
-Instruction* getAndDependence(Value* CurrentAddress) {
-  // If 'CurrentAddress' is tainted, get the OR instruction.
-  auto* OrAddress = getOrAddress(CurrentAddress);
-  if (OrAddress == nullptr) {
-    return nullptr;
-  }
-
-  // No need to check the operands.
-  auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
-  assert(AndDepInst);
-  return AndDepInst;
-}
-
-// Given a value, if it's a tainted address, this function returns
-// the "dependence value", which is the first operand in the AND instruction.
-// E.g., it returns '%1' given 'CurrentAddress' is '%5'.
-// %0 = load i32, i32* @y, align 4, !tbaa !1
-// %cmp = icmp ne i32 %0, 42  // <== this is like the condition
-// %1 = sext i1 %cmp to i32
-// %2 = ptrtoint i32* @x to i32
-// %3 = and i32 %1, 0
-// %4 = or i32 %3, %2
-// %5 = inttoptr i32 %4 to i32*
-// store i32 1, i32* %5, align 4
-Value* getDependence(Value* CurrentAddress) {
-  auto* AndInst = getAndDependence(CurrentAddress);
-  if (AndInst == nullptr) {
-    return nullptr;
-  }
-  return AndInst->getOperand(0);
-}
-
-// Given an address that has been tainted, returns the only condition it depends
-// on, if any; otherwise, returns nullptr.
-Value* getConditionDependence(Value* Address) {
-  auto* Dep = getDependence(Address);
-  if (Dep == nullptr) {
-    // 'Address' has not been dependence-tainted.
-    return nullptr;
-  }
-
-  Value* Operand = Dep;
-  while (true) {
-    auto* Inst = dyn_cast<Instruction>(Operand);
-    if (Inst == nullptr) {
-      // Non-instruction type does not have condition dependence.
-      return nullptr;
-    }
-    if (Inst->getOpcode() == Instruction::ICmp) {
-      return Inst;
-    } else {
-      if (Inst->getNumOperands() != 1) {
-        return nullptr;
-      } else {
-        Operand = Inst->getOperand(0);
-      }
-    }
-  }
-}
-
-// Conservatively decides whether the dependence set of 'Val1' includes the
-// dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
-// 'Val2' and use that single value as its dependence set.
-// If it returns true, it means the dependence set of 'Val1' includes that of
-// 'Val2'; otherwise, it only means we cannot conclusively decide it.
-bool dependenceSetInclusion(Value* Val1, Value* Val2,
-                            int Val1ExpandLevel = 2 * kDependenceDepth,
-                            int Val2ExpandLevel = kDependenceDepth) {
-  typedef SmallSet<Value*, 8> IncludingSet;
-  typedef SmallSet<Value*, 4> IncludedSet;
-
-  IncludingSet DepSet1;
-  IncludedSet DepSet2;
-  // Look for more depths for the including set.
-  recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
-                            Val1ExpandLevel);
-  recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
-                            Val2ExpandLevel);
-
-  auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
-    for (auto* Dep : Subset) {
-      if (0 == FullSet.count(Dep)) {
-        return false;
-      }
-    }
-    return true;
-  };
-  bool inclusion = set_inclusion(DepSet1, DepSet2);
-  DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
-  DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
-  DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
-  DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
-  DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
-
-  return inclusion;
-}
-
-// Recursively iterates through the operands spawned from 'DepVal'. If there
-// exists a single value that 'DepVal' only depends on, we call that value the
-// root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
-Value* getRootDependence(Value* DepVal) {
-  SmallSet<Value*, 8> DepSet;
-  for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
-    recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
-                              depth);
-    if (DepSet.size() == 1) {
-      return *DepSet.begin();
-    }
-    DepSet.clear();
-  }
-  return DepVal;
-}
-
-// This function actually taints 'DepVal' to the address to 'SI'. If the
-// address
-// of 'SI' already depends on whatever 'DepVal' depends on, this function
-// doesn't do anything and returns false. Otherwise, returns true.
-//
-// This effect forces the store and any stores that comes later to depend on
-// 'DepVal'. For example, we have a condition "cond", and a store instruction
-// "s: STORE addr, val". If we want "s" (and any later store) to depend on
-// "cond", we do the following:
-// %conv = sext i1 %cond to i32
-// %addrVal = ptrtoint i32* %addr to i32
-// %andCond = and i32 conv, 0;
-// %orAddr = or i32 %andCond, %addrVal;
-// %NewAddr = inttoptr i32 %orAddr to i32*;
-//
-// This is a more concrete example:
-// ------
-// %0 = load i32, i32* @y, align 4, !tbaa !1
-// %cmp = icmp ne i32 %0, 42  // <== this is like the condition
-// %1 = sext i1 %cmp to i32
-// %2 = ptrtoint i32* @x to i32
-// %3 = and i32 %1, 0
-// %4 = or i32 %3, %2
-// %5 = inttoptr i32 %4 to i32*
-// store i32 1, i32* %5, align 4
-bool taintStoreAddress(StoreInst* SI, Value* DepVal) {
-  // Set the insertion point right after the 'DepVal'.
-  Instruction* Inst = nullptr;
-  IRBuilder<true, NoFolder> Builder(SI);
-  BasicBlock* BB = SI->getParent();
-  Value* Address = SI->getPointerOperand();
-  Type* TargetIntegerType =
-      IntegerType::get(Address->getContext(),
-                       BB->getModule()->getDataLayout().getPointerSizeInBits());
-
-  // Does SI's address already depends on whatever 'DepVal' depends on?
-  if (StoreAddressDependOnValue(SI, DepVal)) {
-    return false;
-  }
-
-  // Figure out if there's a root variable 'DepVal' depends on. For example, we
-  // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
-  // to be "%struct* %0" since all other operands are constant.
-  auto* RootVal = getRootDependence(DepVal);
-  auto* RootInst = dyn_cast<Instruction>(RootVal);
-  auto* DepValInst = dyn_cast<Instruction>(DepVal);
-  if (RootInst && DepValInst &&
-      RootInst->getParent() == DepValInst->getParent()) {
-    DepVal = RootVal;
-  }
-
-  // Is this already a dependence-tainted store?
-  Value* OldDep = getDependence(Address);
-  if (OldDep) {
-    // The address of 'SI' has already been tainted.  Just need to absorb the
-    // DepVal to the existing dependence in the address of SI.
-    Instruction* AndDep = getAndDependence(Address);
-    IRBuilder<true, NoFolder> Builder(AndDep);
-    Value* NewDep = nullptr;
-    if (DepVal->getType() == AndDep->getType()) {
-      NewDep = Builder.CreateAnd(OldDep, DepVal);
-    } else {
-      NewDep = Builder.CreateAnd(
-          OldDep, createCast(Builder, DepVal, TargetIntegerType));
-    }
-
-    auto* NewDepInst = dyn_cast<Instruction>(NewDep);
-
-    // Use the new AND instruction as the dependence
-    AndDep->setOperand(0, NewDep);
-    return true;
-  }
-
-  // SI's address has not been tainted. Now taint it with 'DepVal'.
-  Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
-  Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
-  Value* AndDepVal =
-      Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
-  auto AndInst = dyn_cast<Instruction>(AndDepVal);
-  // XXX-comment: The original IR InstCombiner would change our and instruction
-  // to a select and then the back end optimize the condition out.  We attach a
-  // flag to instructions and set it here to inform the InstCombiner to not to
-  // touch this and instruction at all.
-  Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
-  Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
-
-  DEBUG(dbgs() << "[taintStoreAddress]\n"
-               << "Original store: " << *SI << '\n');
-  SI->setOperand(1, NewAddr);
-
-  // Debug output.
-  DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
-               << "\tCast dependence value to integer: " << *CastDepToInt
-               << '\n'
-               << "\tCast address to integer: " << *PtrToIntCast << '\n'
-               << "\tAnd dependence value: " << *AndDepVal << '\n'
-               << "\tOr address: " << *OrAddr << '\n'
-               << "\tCast or instruction to address: " << *NewAddr << "\n\n");
-
-  return true;
-}
-
-// Looks for the previous store in the if block --- 'BrBB', which makes the
-// speculative store 'StoreToHoist' safe.
-Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
-  assert(StoreToHoist && "StoreToHoist must be a real store");
-
-  Value* StorePtr = StoreToHoist->getPointerOperand();
-
-  // Look for a store to the same pointer in BrBB.
-  for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
-       RI != RE; ++RI) {
-    Instruction* CurI = &*RI;
-
-    StoreInst* SI = dyn_cast<StoreInst>(CurI);
-    // Found the previous store make sure it stores to the same location.
-    // XXX-update: If the previous store's original untainted address are the
-    // same as 'StorePtr', we are also good to hoist the store.
-    if (SI && (SI->getPointerOperand() == StorePtr ||
-               GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
-      // Found the previous store, return its value operand.
-      return SI;
-    }
-  }
-
-  assert(false &&
-         "We should not reach here since this store is safe to speculate");
-}
-
-// XXX-comment: Returns true if it changes the code, false otherwise (the branch
-// condition already depends on 'DepVal'.
-bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
-  assert(BI->isConditional());
-  auto* Cond = BI->getOperand(0);
-  if (dependenceSetInclusion(Cond, DepVal)) {
-    // The dependence/ordering is self-evident.
-    return false;
-  }
-
-  IRBuilder<true, NoFolder> Builder(BI);
-  auto* AndDep =
-      Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
-  auto* TruncAndDep =
-      Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
-  auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
-  BI->setOperand(0, OrCond);
-
-  // Debug output.
-  DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
-
-  return true;
-}
-
-bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
-  assert(BI->isConditional());
-  auto* Cond = BI->getOperand(0);
-  return dependenceSetInclusion(Cond, DepVal);
-}
-
-// XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
-// the first conditional branch. Returns nullptr if there's no such immediately
-// following store/branch instructions, which we can only enforce the load with
-// 'acquire'. 'ChainedBB' contains all the blocks chained together with
-// unconditional branches from 'BB' to the block with the first store/cond
-// branch.
-template <typename Vector>
-Instruction* findFirstStoreCondBranchInst(LoadInst* LI, Vector* ChainedBB) {
-  // In some situations, relaxed loads can be left as is:
-  // 1. The relaxed load is used to calculate the address of the immediate
-  // following store;
-  // 2. The relaxed load is used as a condition in the immediate following
-  // condition, and there are no stores in between. This is actually quite
-  // common. E.g.,
-  // int r1 = x.load(relaxed);
-  // if (r1 != 0) {
-  //   y.store(1, relaxed);
-  // }
-  // However, in this function, we don't deal with them directly. Instead, we
-  // just find the immediate following store/condition branch and return it.
-
-  assert(ChainedBB != nullptr && "Chained BB should not be nullptr");
-  auto* BB = LI->getParent();
-  ChainedBB->push_back(BB);
-  auto BE = BB->end();
-  auto BBI = BasicBlock::iterator(LI);
-  BBI++;
-  while (true) {
-    for (; BBI != BE; BBI++) {
-      Instruction* Inst = &*BBI;
-      IntrinsicInst* II = dyn_cast<IntrinsicInst>(&*BBI);
-      if (II && II->getIntrinsicID() == Intrinsic::aarch64_stlxr) {
-        return II;
-      } else if (Inst->getOpcode() == Instruction::Store) {
-        return Inst;
-      } else if (Inst->getOpcode() == Instruction::Br) {
-        auto* BrInst = dyn_cast<BranchInst>(Inst);
-        if (BrInst->isConditional()) {
-          return Inst;
-        } else {
-          // Reinitialize iterators with the destination of the unconditional
-          // branch.
-          BB = BrInst->getSuccessor(0);
-          ChainedBB->push_back(BB);
-          BBI = BB->begin();
-          BE = BB->end();
-          break;
-        }
-      }
-    }
-    if (BBI == BE) {
-      return nullptr;
-    }
-  }
-}
-
-// XXX-update: Find the next node of the last relaxed load from 'FromInst' to
-// 'ToInst'. If none, return 'ToInst'.
-Instruction* findLastLoadNext(Instruction* FromInst, Instruction* ToInst) {
-  if (FromInst == ToInst) {
-    return ToInst;
-  }
-  Instruction* LastLoad = ToInst;
-  auto* BB = FromInst->getParent();
-  auto BE = BB->end();
-  auto BBI = BasicBlock::iterator(FromInst);
-  BBI++;
-  for (; BBI != BE && &*BBI != ToInst; BBI++) {
-    auto* LI = dyn_cast<LoadInst>(&*BBI);
-    if (LI == nullptr || !LI->isAtomic() || LI->getOrdering() != Monotonic) {
-      continue;
-    }
-    LastLoad = LI;
-    LastLoad = LastLoad->getNextNode();
-  }
-  return LastLoad;
-}
-
-// Inserts a fake conditional branch right after the instruction 'SplitInst',
-// and the branch condition is 'Condition'. 'SplitInst' will be placed in the
-// newly created block.
-void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
-  auto* BB = SplitInst->getParent();
-  TerminatorInst* ThenTerm = nullptr;
-  TerminatorInst* ElseTerm = nullptr;
-  SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
-  assert(ThenTerm && ElseTerm &&
-         "Then/Else terminators cannot be empty after basic block spliting");
-  auto* ThenBB = ThenTerm->getParent();
-  auto* ElseBB = ElseTerm->getParent();
-  auto* TailBB = ThenBB->getSingleSuccessor();
-  assert(TailBB && "Tail block cannot be empty after basic block spliting");
-
-  ThenBB->disableCanEliminateBlock();
-  ThenBB->disableCanEliminateBlock();
-  TailBB->disableCanEliminateBlock();
-  ThenBB->setName(BB->getName() + "Then.Fake");
-  ElseBB->setName(BB->getName() + "Else.Fake");
-  DEBUG(dbgs() << "Add fake conditional branch:\n"
-               << "Then Block:\n"
-               << *ThenBB << "Else Block:\n"
-               << *ElseBB << "\n");
-}
-
-// Returns true if the code is changed, and false otherwise.
-void TaintRelaxedLoads(Instruction* UsageInst, Instruction* InsertPoint) {
-  // For better performance, we can add a "AND X 0" instruction before the
-  // condition.
-  auto* BB = UsageInst->getParent();
-  if (InsertPoint == nullptr) {
-    InsertPoint = UsageInst->getNextNode();
-  }
-  // Insert instructions after PHI nodes.
-  while (dyn_cast<PHINode>(InsertPoint)) {
-    InsertPoint = InsertPoint->getNextNode();
-  }
-  // First thing is to cast 'UsageInst' to an integer type if necessary.
-  Value* AndTarget = nullptr;
-  Type* TargetIntegerType =
-      IntegerType::get(UsageInst->getContext(),
-                       BB->getModule()->getDataLayout().getPointerSizeInBits());
-
-  // Check whether InsertPoint is a added fake conditional branch.
-  BranchInst* BI = nullptr;
-  if ((BI = dyn_cast<BranchInst>(InsertPoint)) && BI->isConditional()) {
-    auto* Cond = dyn_cast<Instruction>(BI->getOperand(0));
-    if (Cond && Cond->getOpcode() == Instruction::ICmp) {
-      auto* CmpInst = dyn_cast<ICmpInst>(Cond);
-      auto* Op0 = dyn_cast<Instruction>(Cond->getOperand(0));
-      auto* Op1 = dyn_cast<ConstantInt>(Cond->getOperand(1));
-      // %tmp = And X, 0
-      // %cmp = ICMP_NE %tmp, 0
-      // Br %cmp
-      // =>
-      // %tmp1 = And X, NewTaintedVal
-      // %tmp2 = And %tmp1, 0
-      // %cmp = ICMP_NE %tmp2, 0
-      // Br %cmp
-      if (CmpInst && CmpInst->getPredicate() == CmpInst::ICMP_NE && Op0 &&
-          Op0->getOpcode() == Instruction::And && Op1 && Op1->isZero()) {
-        auto* Op01 = dyn_cast<ConstantInt>(Op0->getOperand(1));
-        if (Op01 && Op01->isZero()) {
-          // Now we have a previously added fake cond branch.
-          auto* Op00 = Op0->getOperand(0);
-          IRBuilder<true, NoFolder> Builder(CmpInst);
-          if (Op00->getType() == UsageInst->getType()) {
-            AndTarget = UsageInst;
-          } else {
-            AndTarget = createCast(Builder, UsageInst, Op00->getType());
-          }
-          AndTarget = Builder.CreateAnd(Op00, AndTarget);
-          auto* AndZero = dyn_cast<Instruction>(Builder.CreateAnd(
-              AndTarget, Constant::getNullValue(AndTarget->getType())));
-          CmpInst->setOperand(0, AndZero);
-          return;
-        }
-      }
-    }
-  }
-
-  IRBuilder<true, NoFolder> Builder(InsertPoint);
-  if (IntegerType::classof(UsageInst->getType())) {
-    AndTarget = UsageInst;
-  } else {
-    AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
-  }
-  auto* AndZero = dyn_cast<Instruction>(
-      Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
-  auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
-      CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
-  AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
-}
-
-// XXX-comment: Finds the appropriate Value derived from an atomic load.
-// 'ChainedBB' contains all the blocks chained together with unconditional
-// branches from LI's parent BB to the block with the first store/cond branch.
-// If we don't find any, it means 'LI' is not used at all (which should not
-// happen in practice). We can simply set 'LI' to be acquire just to be safe.
-template <typename Vector>
-Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
-                                           Vector* ChainedBB,
-                                           DominatorTree* DT) {
-  typedef SmallSet<Instruction*, 8> UsageSet;
-  typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
-  assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
-  // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
-  // 'LI' in each block.
-  UsageMap usage_map;
-  auto* LoadBB = LI->getParent();
-  usage_map[LoadBB] = make_unique<UsageSet>();
-  usage_map[LoadBB]->insert(LI);
-
-  for (auto* BB : *ChainedBB) {
-    if (usage_map[BB] == nullptr) {
-      usage_map[BB] = make_unique<UsageSet>();
-    }
-    auto& usage_set = usage_map[BB];
-    if (usage_set->size() == 0) {
-      // The value has not been used.
-      return nullptr;
-    }
-    // Calculate the usage in the current BB first.
-    std::list<Value*> bb_usage_list;
-    std::copy(usage_set->begin(), usage_set->end(),
-              std::back_inserter(bb_usage_list));
-    for (auto list_iter = bb_usage_list.begin();
-         list_iter != bb_usage_list.end(); list_iter++) {
-      auto* val = *list_iter;
-      for (auto* U : val->users()) {
-        Instruction* Inst = nullptr;
-        if (!(Inst = dyn_cast<Instruction>(U))) {
-          continue;
-        }
-        assert(Inst && "Usage value must be an instruction");
-        auto iter =
-            std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
-        if (iter == ChainedBB->end()) {
-          // Only care about usage within ChainedBB.
-          continue;
-        }
-        auto* UsageBB = *iter;
-        if (UsageBB == BB) {
-          // Current BB.
-          if (!usage_set->count(Inst)) {
-            bb_usage_list.push_back(Inst);
-            usage_set->insert(Inst);
-          }
-        } else {
-          // A following BB.
-          if (usage_map[UsageBB] == nullptr) {
-            usage_map[UsageBB] = make_unique<UsageSet>();
-          }
-          usage_map[UsageBB]->insert(Inst);
-        }
-      }
-    }
-  }
-
-  // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
-  auto* LaterBB = LaterInst->getParent();
-  auto& usage_set = usage_map[LaterBB];
-  Instruction* usage_inst = nullptr;
-  for (auto* inst : *usage_set) {
-    if (DT->dominates(inst, LaterInst)) {
-      usage_inst = inst;
-      break;
-    }
-  }
-
-  assert(usage_inst && "The usage instruction in the same block but after the "
-                       "later instruction");
-  return usage_inst;
-}
-
-// XXX-comment: Returns whether the code has been changed.
-bool AddFakeConditionalBranchAfterMonotonicLoads(
-    SmallSet<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
-  bool Changed = false;
-  while (!MonotonicLoadInsts.empty()) {
-    auto* LI = *MonotonicLoadInsts.begin();
-    MonotonicLoadInsts.erase(LI);
-    SmallVector<BasicBlock*, 2> ChainedBB;
-    auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
-    if (FirstInst != nullptr) {
-      if (FirstInst->getOpcode() == Instruction::Store) {
-        if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
-          continue;
-        }
-      } else if (FirstInst->getOpcode() == Instruction::Br) {
-        if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
-                                            LI)) {
-          continue;
-        }
-      } else {
-        IntrinsicInst* II = dyn_cast<IntrinsicInst>(FirstInst);
-        if (!II || II->getIntrinsicID() != Intrinsic::aarch64_stlxr) {
-          dbgs() << "FirstInst=" << *FirstInst << "\n";
-          assert(false && "findFirstStoreCondBranchInst() should return a "
-                          "store/condition branch instruction");
-        }
-      }
-    }
-
-    // We really need to process the relaxed load now.
-    StoreInst* SI = nullptr;
-    IntrinsicInst* II = nullptr;
-    if (FirstInst) {
-      SI = dyn_cast<StoreInst>(FirstInst);
-      II = dyn_cast<IntrinsicInst>(FirstInst);
-    }
-    if (FirstInst &&
-        (SI || (II && II->getIntrinsicID() == Intrinsic::aarch64_stlxr))) {
-      // For immediately coming stores, taint the address of the store.
-      if (FirstInst->getParent() == LI->getParent() ||
-          DT->dominates(LI, FirstInst)) {
-        TaintRelaxedLoads(LI, FirstInst);
-        Changed = true;
-      } else {
-        auto* Inst =
-            findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
-        if (!Inst) {
-          LI->setOrdering(Acquire);
-          Changed = true;
-        } else {
-          TaintRelaxedLoads(Inst, FirstInst);
-          Changed = true;
-        }
-      }
-    } else {
-      // No upcoming branch
-      if (!FirstInst) {
-        TaintRelaxedLoads(LI, nullptr);
-        Changed = true;
-      } else {
-        // For immediately coming branch, directly add a fake branch.
-        if (FirstInst->getParent() == LI->getParent() ||
-            DT->dominates(LI, FirstInst)) {
-          TaintRelaxedLoads(LI, FirstInst);
-          Changed = true;
-        } else {
-          auto* Inst =
-              findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
-          if (Inst) {
-            TaintRelaxedLoads(Inst, FirstInst);
-          } else {
-            LI->setOrdering(Acquire);
-          }
-          Changed = true;
-        }
-      }
-    }
-  }
-  return Changed;
-}
-
-// XXX-comment: Returns whether the code has been changed.
-bool ConvertMonotonicLoadsToAcquire(SmallSet<LoadInst*, 1>& MonotonicLoadInsts)
-{
-  bool Changed = false;
-  while (!MonotonicLoadInsts.empty()) {
-    auto* LI = *MonotonicLoadInsts.begin();
-    MonotonicLoadInsts.erase(LI);
-    SmallVector<BasicBlock*, 2> ChainedBB;
-    auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
-    if (FirstInst != nullptr) {
-      if (FirstInst->getOpcode() == Instruction::Store) {
-        if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
-          continue;
-        }
-      } else if (FirstInst->getOpcode() == Instruction::Br) {
-        if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
-                                            LI)) {
-          continue;
-        }
-      } else {
-        IntrinsicInst* II = dyn_cast<IntrinsicInst>(FirstInst);
-        if (!II || II->getIntrinsicID() != Intrinsic::aarch64_stlxr) {
-          dbgs() << "FirstInst=" << *FirstInst << "\n";
-          assert(false && "findFirstStoreCondBranchInst() should return a "
-                          "store/condition branch instruction");
-        }
-      }
-    }
-
-    // We really need convert the load to acquire.
-    LI->setOrdering(Acquire);
-  }
-  return Changed;
-}
-
-/**** Implementations of public methods for dependence tainting ****/
-Value* GetUntaintedAddress(Value* CurrentAddress) {
-  auto* OrAddress = getOrAddress(CurrentAddress);
-  if (OrAddress == nullptr) {
-    // Is it tainted by a select instruction?
-    auto* Inst = dyn_cast<Instruction>(CurrentAddress);
-    if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
-      // A selection instruction.
-      if (Inst->getOperand(1) == Inst->getOperand(2)) {
-        return Inst->getOperand(1);
-      }
-    }
-
-    return CurrentAddress;
-  }
-  Value* ActualAddress = nullptr;
-
-  auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
-  if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
-    return CastToInt->getOperand(0);
-  } else {
-    // This should be a IntToPtr constant expression.
-    ConstantExpr* PtrToIntExpr =
-        dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
-    if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
-      return PtrToIntExpr->getOperand(0);
-    }
-  }
-
-  // Looks like it's not been dependence-tainted. Returns itself.
-  return CurrentAddress;
-}
-
-MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
-  AAMDNodes AATags;
-  SI->getAAMetadata(AATags);
-  const auto& DL = SI->getModule()->getDataLayout();
-  const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
-  DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
-    dbgs() << "[GetUntaintedMemoryLocation]\n"
-           << "Storing address: " << *SI->getPointerOperand()
-           << "\nUntainted address: " << *OriginalAddr << "\n";
-  });
-  return MemoryLocation(OriginalAddr,
-                        DL.getTypeStoreSize(SI->getValueOperand()->getType()),
-                        AATags);
-}
-
-bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
-  if (dependenceSetInclusion(SI, DepVal)) {
-    return false;
-  }
-
-  bool tainted = taintStoreAddress(SI, DepVal);
-  assert(tainted);
-  return tainted;
-}
-
-bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
-  if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
-    return false;
-  }
-
-  bool tainted = taintStoreAddress(SI, DepVal);
-  assert(tainted);
-  return tainted;
-}
-
-bool CompressTaintedStore(BasicBlock* BB) {
-  // This function looks for windows of adajcent stores in 'BB' that satisfy the
-  // following condition (and then do optimization):
-  // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
-  //                 address depends on && Dep(v1) includes Dep(d1);
-  // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
-  //                 address depends on && Dep(v2) includes Dep(d2) &&
-  //                 Dep(d2) includes Dep(d1);
-  // ...
-  // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
-  //                 address depends on && Dep(dN) includes Dep(d"N-1").
-  //
-  // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
-  // safely transform the above to the following. In between these stores, we
-  // can omit untainted stores to the same address 'Addr' since they internally
-  // have dependence on the previous stores on the same address.
-  // =>
-  // *Addr = v1
-  // *Addr = v2
-  // *Addr(d3) = v3
-  for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
-    // Look for the first store in such a window of adajacent stores.
-    auto* FirstSI = dyn_cast<StoreInst>(&*BI);
-    if (!FirstSI) {
-      continue;
-    }
-
-    // The first store in the window must be tainted.
-    auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
-    if (UntaintedAddress == FirstSI->getPointerOperand()) {
-      continue;
-    }
-
-    // The first store's address must directly depend on and only depend on a
-    // condition.
-    auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
-    if (nullptr == FirstSIDepCond) {
-      continue;
-    }
-
-    // Dep(first store's storing value) includes Dep(tainted dependence).
-    if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
-      continue;
-    }
-
-    // Look for subsequent stores to the same address that satisfy the condition
-    // of "compressing the dependence".
-    SmallVector<StoreInst*, 8> AdajacentStores;
-    AdajacentStores.push_back(FirstSI);
-    auto BII = BasicBlock::iterator(FirstSI);
-    for (BII++; BII != BE; BII++) {
-      auto* CurrSI = dyn_cast<StoreInst>(&*BII);
-      if (!CurrSI) {
-        if (BII->mayHaveSideEffects()) {
-          // Be conservative. Instructions with side effects are similar to
-          // stores.
-          break;
-        }
-        continue;
-      }
-
-      auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
-      auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
-      // All other stores must satisfy either:
-      // A. 'CurrSI' is an untainted store to the same address, or
-      // B. the combination of the following 5 subconditions:
-      // 1. Tainted;
-      // 2. Untainted address is the same as the group's address;
-      // 3. The address is tainted with a sole value which is a condition;
-      // 4. The storing value depends on the condition in 3.
-      // 5. The condition in 3 depends on the previous stores dependence
-      // condition.
-
-      // Condition A. Should ignore this store directly.
-      if (OrigAddress == CurrSI->getPointerOperand() &&
-          OrigAddress == UntaintedAddress) {
-        continue;
-      }
-      // Check condition B.
-      Value* Cond = nullptr;
-      if (OrigAddress == CurrSI->getPointerOperand() ||
-          OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
-          !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
-        // Check condition 1, 2, 3 & 4.
-        break;
-      }
-
-      // Check condition 5.
-      StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
-      auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
-      assert(PrevSIDepCond &&
-             "Store in the group must already depend on a condtion");
-      if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
-        break;
-      }
-
-      AdajacentStores.push_back(CurrSI);
-    }
-
-    if (AdajacentStores.size() == 1) {
-      // The outer loop should keep looking from the next store.
-      continue;
-    }
-
-    // Now we have such a group of tainted stores to the same address.
-    DEBUG(dbgs() << "[CompressTaintedStore]\n");
-    DEBUG(dbgs() << "Original BB\n");
-    DEBUG(dbgs() << *BB << '\n');
-    auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
-    for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
-      auto* SI = AdajacentStores[i];
-
-      // Use the original address for stores before the last one.
-      SI->setOperand(1, UntaintedAddress);
-
-      DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
-    }
-    // XXX-comment: Try to make the last store use fewer registers.
-    // If LastSI's storing value is a select based on the condition with which
-    // its address is tainted, transform the tainted address to a select
-    // instruction, as follows:
-    // r1 = Select Cond ? A : B
-    // r2 = Cond & 0
-    // r3 = Addr | r2
-    // *r3 = r1
-    // ==>
-    // r1 = Select Cond ? A : B
-    // r2 = Select Cond ? Addr : Addr
-    // *r2 = r1
-    // The idea is that both Select instructions depend on the same condition,
-    // so hopefully the backend can generate two cmov instructions for them (and
-    // this saves the number of registers needed).
-    auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
-    auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
-    if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
-        LastSIValue->getOperand(0) == LastSIDep) {
-      // XXX-comment: Maybe it's better for us to just leave it as an and/or
-      // dependence pattern.
-      //      /*
-      IRBuilder<true, NoFolder> Builder(LastSI);
-      auto* Address =
-          Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
-      LastSI->setOperand(1, Address);
-      DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
-      //      */
-    }
-  }
-
-  return true;
-}
-
-bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
-  Value* OldDep = getDependence(OldAddress);
-  // Return false when there's no dependence to pass from the OldAddress.
-  if (!OldDep) {
-    return false;
-  }
-
-  // No need to pass the dependence to NewStore's address if it already depends
-  // on whatever 'OldAddress' depends on.
-  if (StoreAddressDependOnValue(NewStore, OldDep)) {
-    return false;
-  }
-  return taintStoreAddress(NewStore, OldAddress);
-}
-
-SmallSet<Value*, 8> FindDependence(Value* Val) {
-  SmallSet<Value*, 8> DepSet;
-  recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
-  return DepSet;
-}
-
-bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
-  return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
-}
-
-bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
-  return dependenceSetInclusion(SI, Dep);
-}
-
-} // namespace
-
-
-
 bool CodeGenPrepare::runOnFunction(Function &F) {
-  bool EverMadeChange = false;
-
   if (skipOptnoneFunction(F))
     return false;
 
   DL = &F.getParent()->getDataLayout();
 
+  bool EverMadeChange = false;
   // Clear per function information.
   InsertedInsts.clear();
   PromotedInsts.clear();
@@ -1320,7 +218,6 @@ bool CodeGenPrepare::runOnFunction(Function &F) {
     TLI = TM->getSubtargetImpl(F)->getTargetLowering();
   TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
-  DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   OptSize = F.optForSize();
 
   /// This optimization identifies DIV instructions that can be
@@ -1419,6 +316,7 @@ bool CodeGenPrepare::runOnFunction(Function &F) {
     for (auto &I : Statepoints)
       EverMadeChange |= simplifyOffsetableRelocate(*I);
   }
+
   return EverMadeChange;
 }
 
@@ -1465,6 +363,7 @@ bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
   // Note that this intentionally skips the entry block.
   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
     BasicBlock *BB = &*I++;
+
     // If this block doesn't end with an uncond branch, ignore it.
     BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
     if (!BI || !BI->isUnconditional())
@@ -1512,7 +411,7 @@ bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
       const Instruction *UI = cast<Instruction>(U);
       if (UI->getParent() != DestBB || !isa<PHINode>(UI))
         return false;
-      // IfUser is inside DestBB block and it is a PHINode then check
+      // If User is inside DestBB block and it is a PHINode then check
       // incoming value. If incoming value is not from BB then this is
       // a complex condition (e.g. preheaders) we want to avoid here.
       if (UI->getParent() == DestBB) {
diff --git a/lib/CodeGen/MachineBasicBlock.cpp b/lib/CodeGen/MachineBasicBlock.cpp
index 42190168742..85d544d9498 100644
--- a/lib/CodeGen/MachineBasicBlock.cpp
+++ b/lib/CodeGen/MachineBasicBlock.cpp
@@ -40,7 +40,7 @@ using namespace llvm;
 #define DEBUG_TYPE "codegen"
 
 MachineBasicBlock::MachineBasicBlock(MachineFunction &MF, const BasicBlock *B)
-    : BB(B), Number(-1), xParent(&MF), canEliminateMachineBB(true) {
+    : BB(B), Number(-1), xParent(&MF) {
   Insts.Parent = this;
 }
 
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index f119023d217..c741982bc08 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -924,62 +924,6 @@ CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
   TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
   APInt KnownZero, KnownOne;
-
-  // XXX-disabled:
-  auto Opcode = Op.getOpcode();
-  if (Opcode == ISD::AND || Opcode == ISD::OR) {
-    auto* Op1 = Op.getOperand(0).getNode();
-    auto* Op2 = Op.getOperand(1).getNode();
-    auto* Op1C = dyn_cast<ConstantSDNode>(Op1);
-    auto* Op2C = dyn_cast<ConstantSDNode>(Op2);
-
-    // and X, 0
-    if (Opcode == ISD::AND && !Op1C && Op2C && Op2C->isNullValue()) {
-      return false;
-    }
-
-    // or (and X, 0), Y
-    if (Opcode == ISD::OR) {
-      if (Op1->getOpcode() == ISD::AND) {
-        auto* Op11 = Op1->getOperand(0).getNode();
-        auto* Op12 = Op1->getOperand(1).getNode();
-        auto* Op11C = dyn_cast<ConstantSDNode>(Op11);
-        auto* Op12C = dyn_cast<ConstantSDNode>(Op12);
-        if (!Op11C && Op12C && Op12C->isNullValue()) {
-          return false;
-        }
-      }
-      if (Op1->getOpcode() == ISD::TRUNCATE) {
-        // or (trunc (and %0, 0)), Y
-        auto* Op11 = Op1->getOperand(0).getNode();
-        if (Op11->getOpcode() == ISD::AND) {
-          auto* Op111 = Op11->getOperand(0).getNode();
-          auto* Op112 = Op11->getOperand(1).getNode();
-          auto* Op111C = dyn_cast<ConstantSDNode>(Op111);
-          auto* Op112C = dyn_cast<ConstantSDNode>(Op112);
-          if (!Op111C && Op112C && Op112C->isNullValue()) {
-            // or (and X, 0), Y
-            return false;
-          }
-        }
-      }
-    }
-  }
-
-  // trunc (and X, 0)
-  if (Opcode == ISD::TRUNCATE) {
-    auto* Op1 = Op.getOperand(0).getNode();
-    if (Op1->getOpcode() == ISD::AND) {
-      auto* Op11 = Op1->getOperand(0).getNode();
-      auto* Op12 = Op1->getOperand(1).getNode();
-      auto* Op11C = dyn_cast<ConstantSDNode>(Op11);
-      auto* Op12C = dyn_cast<ConstantSDNode>(Op12);
-      if (!Op11C && Op12C && Op12C->isNullValue()) {
-        return false;
-      }
-    }
-  }
-
   if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
     return false;
 
@@ -3098,22 +3042,6 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
   // fold (and c1, c2) -> c1&c2
   ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
   ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
-
-  // XXX-disabled: (and x, 0) should not be folded.
-  // (and (and x, 0), y) shouldn't either.
-  if (!N0C && N1C && N1C->isNullValue()) {
-    return SDValue();
-  }
-  if (!N0C) {
-    if (N0.getOpcode() == ISD::AND) {
-      auto* N01 = N0.getOperand(1).getNode();
-      auto* N01C = dyn_cast<ConstantSDNode>(N01);
-      if (N01C && N01C->isNullValue()) {
-        return SDValue();
-      }
-    }
-  }
-
   if (N0C && N1C && !N1C->isOpaque())
     return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C);
   // canonicalize constant to RHS
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index e845a9919ef..893871f9448 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -3465,14 +3465,10 @@ SDValue SelectionDAG::getNode(unsigned Opcode, SDLoc DL, EVT VT, SDValue N1,
     assert(VT.isInteger() && "This operator does not apply to FP types!");
     assert(N1.getValueType() == N2.getValueType() &&
            N1.getValueType() == VT && "Binary operator types must match!");
-
-    // XXX-disabled:
-    /*
     // (X & 0) -> 0.  This commonly occurs when legalizing i64 values, so it's
     // worth handling here.
     if (N2C && N2C->isNullValue())
       return N2;
-    */
     if (N2C && N2C->isAllOnesValue())  // X & -1 -> X
       return N1;
     break;
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index 491d7ec451c..c075da4738a 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -645,50 +645,6 @@ bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
   // at this point.
   FuncInfo->clear();
 
-  // XXX-update: Right after instruction selection, check through the
-  // intentionally added fake conditional branches and mark them as unremovable.
-  for (auto& MBB : *MF) {
-    // Check whether MBB has two successors which only contains an unconditional
-    // branch to the same destination.
-    if (MBB.succ_size() != 2 ||
-        !MBB.getLastNonDebugInstr()->isUnconditionalBranch()) {
-      continue;
-    }
-    auto MBBSuccIter = MBB.succ_begin();
-    auto* Succ1 = *MBBSuccIter;
-    MBBSuccIter++;
-    auto* Succ2 = *MBBSuccIter;
-
-    MachineBasicBlock* Succ1Succ = nullptr;
-    MachineBasicBlock* Succ2Succ = nullptr;
-    if ((Succ1->size() == 1 && Succ1->begin()->isUnconditionalBranch()) ||
-        (Succ1->size() == 0)) {
-      if (Succ1->succ_size()) {
-        Succ1Succ = *Succ1->succ_begin();
-      }
-    }
-    if ((Succ2->size() == 1 && Succ2->begin()->isUnconditionalBranch()) ||
-        (Succ2->size() == 0)) {
-      if (Succ1->succ_size()) {
-        Succ2Succ = *Succ2->succ_begin();
-      }
-    }
-
-    bool HasCommonDest = Succ1Succ && Succ1Succ == Succ2Succ;
-    if (HasCommonDest) {
-      auto MBBIter = MBB.end();
-      std::advance(MBBIter, -2);
-      assert(MBBIter->isConditionalBranch());
-      MBB.disableCanEliminateMachineBB();
-      Succ1->disableCanEliminateMachineBB();
-      Succ2->disableCanEliminateMachineBB();
-      Succ1Succ->disableCanEliminateMachineBB();
-      DEBUG(dbgs() << "Mark as unremovable machine basic block: " << MBB
-                   << "\nMark as unremovable branch instruction: " << *MBBIter
-                   << "\n");
-    }
-  }
-
   DEBUG(dbgs() << "*** MachineFunction at end of ISel ***\n");
   DEBUG(MF->print(dbgs()));
 
diff --git a/lib/IR/BasicBlock.cpp b/lib/IR/BasicBlock.cpp
index 8e3cac27f48..f61276fd436 100644
--- a/lib/IR/BasicBlock.cpp
+++ b/lib/IR/BasicBlock.cpp
@@ -40,8 +40,7 @@ template class llvm::SymbolTableListTraits<Instruction>;
 
 BasicBlock::BasicBlock(LLVMContext &C, const Twine &Name, Function *NewParent,
                        BasicBlock *InsertBefore)
-  : Value(Type::getLabelTy(C), Value::BasicBlockVal), Parent(nullptr),
-    canEliminateBlock(true) {
+  : Value(Type::getLabelTy(C), Value::BasicBlockVal), Parent(nullptr) {
 
   if (NewParent)
     insertInto(NewParent, InsertBefore);
diff --git a/lib/IR/Verifier.cpp b/lib/IR/Verifier.cpp
index 37b56f8efc7..9198b0e1fb5 100644
--- a/lib/IR/Verifier.cpp
+++ b/lib/IR/Verifier.cpp
@@ -1974,10 +1974,6 @@ void Verifier::visitBasicBlock(BasicBlock &BB) {
              "PHI nodes must have at least one entry.  If the block is dead, "
              "the PHI should be removed!",
              PN);
-      if (PN->getNumIncomingValues() != Preds.size()) {
-        dbgs() << "Problematic function: \n" << *PN->getParent()->getParent() << "\n";
-        dbgs() << "Problematic block: \n" << *PN->getParent() << "\n";
-      }
       Assert(PN->getNumIncomingValues() == Preds.size(),
              "PHINode should have one entry for each predecessor of its "
              "parent basic block!",
@@ -3358,9 +3354,6 @@ void Verifier::verifyDominatesUse(Instruction &I, unsigned i) {
   }
 
   const Use &U = I.getOperandUse(i);
-  if (!(InstsInThisBlock.count(Op) || DT.dominates(Op, U))) {
-    dbgs() << "Problematic function: \n" << *I.getParent()->getParent() << "\n";
-  }
   Assert(InstsInThisBlock.count(Op) || DT.dominates(Op, U),
          "Instruction does not dominate all uses!", Op, &I);
 }
diff --git a/lib/Target/AArch64/AArch64TargetMachine.cpp b/lib/Target/AArch64/AArch64TargetMachine.cpp
index 8cc3ff22725..c52c5544fc7 100644
--- a/lib/Target/AArch64/AArch64TargetMachine.cpp
+++ b/lib/Target/AArch64/AArch64TargetMachine.cpp
@@ -216,9 +216,6 @@ void AArch64PassConfig::addIRPasses() {
   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
   // ourselves.
   addPass(createAtomicExpandPass(TM));
-  // XXX-update: Immediate add -licm pass after atomic expand pass to deal with
-  // loop invariants introduced mannually.
-//  addPass(createLICMPass());
 
   // Cmpxchg instructions are often used with a subsequent comparison to
   // determine whether it succeeded. We can exploit existing control-flow in