diff --git a/include/swift/SILAnalysis/ARCAnalysis.h b/include/swift/SILAnalysis/ARCAnalysis.h
index 0c20bed9e25..82a44130f7e 100644
--- a/include/swift/SILAnalysis/ARCAnalysis.h
+++ b/include/swift/SILAnalysis/ARCAnalysis.h
@@ -15,6 +15,7 @@
 
 #include "swift/SIL/SILValue.h"
 #include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SetVector.h"
 
 namespace swift {
 
@@ -39,10 +40,10 @@ bool canUseValue(SILInstruction *User, SILValue Ptr, AliasAnalysis *AA);
 /// insertion pts, and decrement insertion pts.
 struct ARCMatchingSet {
   SILValue Ptr;
-  llvm::SmallPtrSet<SILInstruction *, 8> Increments;
-  llvm::SmallPtrSet<SILInstruction *, 8> IncrementInsertPts;
-  llvm::SmallPtrSet<SILInstruction *, 8> Decrements;
-  llvm::SmallPtrSet<SILInstruction *, 8> DecrementInsertPts;
+  llvm::SetVector<SILInstruction *> Increments;
+  llvm::SetVector<SILInstruction *> IncrementInsertPts;
+  llvm::SetVector<SILInstruction *> Decrements;
+  llvm::SetVector<SILInstruction *> DecrementInsertPts;
 
   void clear() {
     Ptr = SILValue();
diff --git a/lib/SILAnalysis/GlobalARCSequenceDataflow.cpp b/lib/SILAnalysis/GlobalARCSequenceDataflow.cpp
index 9a0bfa94bfc..6ca0c8743ef 100644
--- a/lib/SILAnalysis/GlobalARCSequenceDataflow.cpp
+++ b/lib/SILAnalysis/GlobalARCSequenceDataflow.cpp
@@ -39,9 +39,9 @@ static bool isAutoreleasePoolCall(SILInstruction &I) {
     return false;
 
   return llvm::StringSwitch<bool>(FRI->getReferencedFunction()->getName())
-    .Case("objc_autoreleasePoolPush", true)
-    .Case("objc_autoreleasePoolPop", true)
-    .Default(false);
+      .Case("objc_autoreleasePoolPush", true)
+      .Case("objc_autoreleasePoolPop", true)
+      .Default(false);
 }
 
 namespace llvm {
@@ -61,7 +61,7 @@ raw_ostream &operator<<(raw_ostream &OS,
 }
 
 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
-                        TopDownRefCountState::LatticeState S) {
+                              TopDownRefCountState::LatticeState S) {
   using LatticeState = TopDownRefCountState::LatticeState;
   switch (S) {
   case LatticeState::None:
@@ -76,41 +76,216 @@ llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
 }
 } // end namespace llvm
 
+//===----------------------------------------------------------------------===//
+//                           Lattice State Merging
+//===----------------------------------------------------------------------===//
+
+static inline BottomUpRefCountState::LatticeState
+MergeBottomUpLatticeStates(BottomUpRefCountState::LatticeState L1,
+                           BottomUpRefCountState::LatticeState L2) {
+  using LatticeState = BottomUpRefCountState::LatticeState;
+  // If both states equal, return the first.
+  if (L1 == L2)
+    return L1;
+
+  // If either are none, return None.
+  if (L1 == LatticeState::None || L2 == LatticeState::None)
+    return LatticeState::None;
+
+  // Canonicalize.
+  if (unsigned(L1) > unsigned(L2))
+    std::swap(L1, L2);
+
+  // Choose the side further along in the sequence.
+  if ((L1 == LatticeState::Decremented || L1 == LatticeState::MightBeUsed) ||
+      (L2 == LatticeState::MightBeUsed ||
+       L2 == LatticeState::MightBeDecremented))
+    return L2;
+
+  // Otherwise, we don't know what happened, be conservative and return none.
+  return LatticeState::None;
+}
+
+static inline TopDownRefCountState::LatticeState
+MergeTopDownLatticeStates(TopDownRefCountState::LatticeState L1,
+                          TopDownRefCountState::LatticeState L2) {
+  using LatticeState = TopDownRefCountState::LatticeState;
+  // If both states equal, return the first.
+  if (L1 == L2)
+    return L1;
+
+  // If either are none, return None.
+  if (L1 == LatticeState::None || L2 == LatticeState::None)
+    return LatticeState::None;
+
+  // Canonicalize.
+  if (unsigned(L1) > unsigned(L2))
+    std::swap(L1, L2);
+
+  // Choose the side further along in the sequence.
+  if ((L1 == LatticeState::Incremented ||
+       L1 == LatticeState::MightBeDecremented) ||
+      (L2 == LatticeState::MightBeDecremented ||
+       L2 == LatticeState::MightBeUsed))
+    return L2;
+
+  // Otherwise, we don't know what happened, return none.
+  return LatticeState::None;
+}
+
 //===----------------------------------------------------------------------===//
 //                         ARCBBState Implementation
 //===----------------------------------------------------------------------===//
 
 /// Merge in the state of the successor basic block. This is currently a stub.
 void ARCBBState::mergeSuccBottomUp(ARCBBState &SuccBB) {
+  // For each entry in the other set, if our set has an entry with the same key,
+  // merge the entires. Otherwise, copy the entry and merge it with an empty
+  // entry.
+  for (auto MI : SuccBB.getBottomupStates()) {
+    auto Pair = PtrToBottomUpState.insert(MI);
+    // If we fail to merge, bail.
+    if (!Pair.first->second.merge(Pair.second ? BottomUpRefCountState()
+                                              : MI.second)) {
+      clear();
+      return;
+    }
+  }
 
+  for (auto Pair : getBottomupStates()) {
+    if (SuccBB.PtrToBottomUpState.find(Pair.first) ==
+        SuccBB.PtrToBottomUpState.end())
+      // If we fail to merge, bail.
+      if (!Pair.second.merge(BottomUpRefCountState())) {
+        clear();
+        return;
+      }
+  }
 }
 
 /// Initialize this BB with the state of the successor basic block. This is
 /// called on a basic block's state and then any other successors states are
 /// merged in. This is currently a stub.
 void ARCBBState::initSuccBottomUp(ARCBBState &SuccBB) {
-
+  PtrToBottomUpState = SuccBB.PtrToBottomUpState;
 }
 
 /// Merge in the state of the predecessor basic block. This is currently a stub.
 void ARCBBState::mergePredTopDown(ARCBBState &PredBB) {
+  // For each entry in the other set, if our set has an entry with the same key,
+  // merge the entires. Otherwise, copy the entry and merge it with an empty
+  // entry.
+  for (auto MI : PredBB.getTopDownStates()) {
+    auto Pair = PtrToTopDownState.insert(MI);
+    // If we fail to merge, bail.
+    if (!Pair.first->second.merge(Pair.second ? TopDownRefCountState()
+                                              : MI.second)) {
+      clear();
+      return;
+    }
+  }
 
+  for (auto Pair : getTopDownStates()) {
+    if (PredBB.PtrToTopDownState.find(Pair.first) ==
+        PredBB.PtrToTopDownState.end())
+      // If we fail to merge, bail.
+      if (!Pair.second.merge(TopDownRefCountState())) {
+        clear();
+        return;
+      }
+  }
 }
 
 /// Initialize the state for this BB with the state of its predecessor
 /// BB. Used to create an initial state before we merge in other
 /// predecessors. This is currently a stub.
 void ARCBBState::initPredTopDown(ARCBBState &PredBB) {
-
+  PtrToTopDownState = PredBB.PtrToTopDownState;
 }
 
 //===----------------------------------------------------------------------===//
 //                    Reference Count State Implementation
 //===----------------------------------------------------------------------===//
 
-void TopDownRefCountState::merge(const TopDownRefCountState &Other) {}
+bool TopDownRefCountState::merge(const TopDownRefCountState &Other) {
+  auto NewState = MergeTopDownLatticeStates(LatState, Other.LatState);
+  DEBUG(llvm::dbgs() << "            Performing TopDown Merge.\n");
+  DEBUG(llvm::dbgs() << "                Left: " << LatState << "; Right: "
+                     << Other.LatState << "; Result: " << NewState << "\n");
+  LatState = NewState;
+  KnownSafe &= KnownSafe;
 
-void BottomUpRefCountState::merge(const BottomUpRefCountState &Other) {}
+  // If we're doing a merge on a path that's previously seen a partial merge,
+  // conservatively drop the sequence, to avoid doing partial RR
+  // elimination. If the branch predicates for the two merge differ, mixing
+  // them is unsafe since they are not control dependent.
+  if (LatState == TopDownRefCountState::LatticeState::None || Partial ||
+      Other.Partial) {
+    RefCountState<TopDownRefCountState>::clear();
+    DEBUG(llvm::dbgs() << "            Found LatticeState::None or Partial. "
+                          "Clearing State!\n");
+    return false;
+  }
+
+  // We should never have an argument path merge with a non-argument path.
+  if (!Argument.isNull()) {
+    RefCountState<TopDownRefCountState>::clear();
+    DEBUG(
+        llvm::dbgs() << "            Can not merge Argument with Non-Argument "
+                        "path... Bailing!\n");
+    return false;
+  }
+
+  Increments.insert(Other.Increments.begin(), Other.Increments.end());
+
+  Partial = InsertPts.size() != Other.InsertPts.size();
+  for (auto *SI : Other.InsertPts)
+    Partial |= InsertPts.insert(SI);
+
+  if (Partial) {
+    DEBUG(llvm::dbgs() << "            Found partial, clearing state!\n");
+    RefCountState<TopDownRefCountState>::clear();
+    return false;
+  }
+
+  return true;
+}
+
+bool BottomUpRefCountState::merge(const BottomUpRefCountState &Other) {
+
+  auto NewState = MergeBottomUpLatticeStates(LatState, Other.LatState);
+  DEBUG(llvm::dbgs() << "            Performing BottomUp Merge.\n");
+  DEBUG(llvm::dbgs() << "                Left: " << LatState << "; Right: "
+                     << Other.LatState << "; Result: " << NewState << "\n");
+  LatState = NewState;
+  KnownSafe &= KnownSafe;
+
+  // If we're doing a merge on a path that's previously seen a partial merge,
+  // conservatively drop the sequence, to avoid doing partial RR
+  // elimination. If the branch predicates for the two merge differ, mixing
+  // them is unsafe since they are not control dependent.
+  if (LatState == BottomUpRefCountState::LatticeState::None || Partial ||
+      Other.Partial) {
+    DEBUG(llvm::dbgs() << "            Found LatticeState::None or Partial. "
+                          "Clearing State!\n");
+    RefCountState<BottomUpRefCountState>::clear();
+    return false;
+  }
+
+  Decrements.insert(Other.Decrements.begin(), Other.Decrements.end());
+
+  Partial = InsertPts.size() != Other.InsertPts.size();
+  for (auto *SI : Other.InsertPts)
+    Partial |= InsertPts.insert(SI);
+
+  if (Partial) {
+    DEBUG(llvm::dbgs() << "            Found partial, clearing state!\n");
+    RefCountState<BottomUpRefCountState>::clear();
+    return false;
+  }
+
+  return true;
+}
 
 //===----------------------------------------------------------------------===//
 //                             Top Down Dataflow
@@ -122,11 +297,10 @@ void BottomUpRefCountState::merge(const BottomUpRefCountState &Other) {}
 ///
 /// NestingDetected will be set to indicate that the block needs to be
 /// reanalyzed if code motion occurs.
-static bool
-processBBTopDown(ARCBBState &BBState,
-                 BlotMapVector<SILInstruction *,
-                               TopDownRefCountState> &DecToIncStateMap,
-                 AliasAnalysis *AA) {
+static bool processBBTopDown(
+    ARCBBState &BBState,
+    BlotMapVector<SILInstruction *, TopDownRefCountState> &DecToIncStateMap,
+    AliasAnalysis *AA) {
   DEBUG(llvm::dbgs() << ">>>> Top Down!\n");
 
   SILBasicBlock &BB = BBState.getBB();
@@ -140,7 +314,7 @@ processBBTopDown(ARCBBState &BBState,
   if (&BB == &*BB.getParent()->begin()) {
     auto Args = BB.getBBArgs();
     auto SignatureParams =
-      BB.getParent()->getLoweredFunctionType()->getInterfaceParameters();
+        BB.getParent()->getLoweredFunctionType()->getInterfaceParameters();
     for (unsigned i = 0, e = Args.size(); i != e; ++i) {
       SILArgument *A = Args[i];
       ParameterConvention P = SignatureParams[i].getConvention();
@@ -176,7 +350,7 @@ processBBTopDown(ARCBBState &BBState,
       NestingDetected |= State.initWithInst(&I);
 
       DEBUG(llvm::dbgs() << "    REF COUNT INCREMENT! Known Safe: "
-            << (State.isKnownSafe()?"yes":"no") << "\n");
+                         << (State.isKnownSafe() ? "yes" : "no") << "\n");
 
       // Continue processing in case this increment could be a CanUse for a
       // different pointer.
@@ -198,15 +372,15 @@ processBBTopDown(ARCBBState &BBState,
         // Copy the current value of ref count state into the result map.
         DecToIncStateMap[&I] = RefCountState;
         DEBUG(llvm::dbgs() << "    MATCHING INCREMENT:\n"
-              << RefCountState.getValue());
+                           << RefCountState.getValue());
 
         // Clear the ref count state in case we see more operations on this
         // ref counted value. This is for safety reasons.
         RefCountState.clear();
       } else {
         if (RefCountState.isTrackingRefCountInst()) {
-          DEBUG(llvm::dbgs() << "    FAILED MATCH INCREMENT:\n" <<
-                RefCountState.getValue());
+          DEBUG(llvm::dbgs() << "    FAILED MATCH INCREMENT:\n"
+                             << RefCountState.getValue());
         } else {
           DEBUG(llvm::dbgs() << "    FAILED MATCH. NO INCREMENT.\n");
         }
@@ -224,8 +398,8 @@ processBBTopDown(ARCBBState &BBState,
       if (Op && OtherState.first == Op)
         continue;
 
-      // If we are tracking an argument, skip it.
-      if (!OtherState.second.isTrackingRefCountInst())
+      // If the other state is not tracking anything, bail.
+      if (!OtherState.second.isTrackingRefCount())
         continue;
 
       // Check if the instruction we are visiting could potentially decrement
@@ -233,7 +407,7 @@ processBBTopDown(ARCBBState &BBState,
       // cause us to change states. If we do change states continue...
       if (OtherState.second.handlePotentialDecrement(&I, AA)) {
         DEBUG(llvm::dbgs() << "    Found Potential Decrement:\n        "
-              << OtherState.second.getValue());
+                           << OtherState.second.getValue());
         continue;
       }
 
@@ -241,7 +415,7 @@ processBBTopDown(ARCBBState &BBState,
       // could be used by the given instruction.
       if (OtherState.second.handlePotentialUser(&I, AA))
         DEBUG(llvm::dbgs() << "    Found Potential Use:\n        "
-              << OtherState.second.getValue());
+                           << OtherState.second.getValue());
     }
   }
 
@@ -249,8 +423,8 @@ processBBTopDown(ARCBBState &BBState,
 }
 
 void
-swift::arc::ARCSequenceDataflowEvaluator::
-mergePredecessors(ARCBBState &BBState, SILBasicBlock *BB) {
+swift::arc::ARCSequenceDataflowEvaluator::mergePredecessors(ARCBBState &BBState,
+                                                            SILBasicBlock *BB) {
   // For each successor of BB...
   unsigned i = 0;
   for (auto Pred : BB->getPreds()) {
@@ -277,13 +451,18 @@ mergePredecessors(ARCBBState &BBState, SILBasicBlock *BB) {
 bool swift::arc::ARCSequenceDataflowEvaluator::processTopDown() {
   bool NestingDetected = false;
 
+  DEBUG(llvm::dbgs() << "<<<< Processing Bottom Up! >>>>\n");
+
   // For each BB in our reverse post order...
   for (auto *BB : reversed(PostOrder)) {
 
+    DEBUG(llvm::dbgs() << "Processing BB#: " << BBToPostOrderID[BB] << "\n");
+
     // Grab the BBState associated with it and set it to be the current BB.
     ARCBBState &BBState = TopDownBBStates[BB];
     BBState.init(BB);
 
+    DEBUG(llvm::dbgs() << "Merging Predecessors!\n");
     mergePredecessors(BBState, BB);
 
     // Then perform the basic block optimization.
@@ -297,18 +476,16 @@ bool swift::arc::ARCSequenceDataflowEvaluator::processTopDown() {
 //                             Bottom Up Dataflow
 //===----------------------------------------------------------------------===//
 
-
 /// Analyze a single BB for refcount inc/dec instructions.
 ///
 /// If anything was found it will be added to DecToIncStateMap.
 ///
 /// NestingDetected will be set to indicate that the block needs to be
 /// reanalyzed if code motion occurs.
-static bool
-processBBBottomUp(ARCBBState &BBState,
-                  BlotMapVector<SILInstruction *,
-                                  BottomUpRefCountState> &IncToDecStateMap,
-                  AliasAnalysis *AA) {
+static bool processBBBottomUp(
+    ARCBBState &BBState,
+    BlotMapVector<SILInstruction *, BottomUpRefCountState> &IncToDecStateMap,
+    AliasAnalysis *AA) {
   DEBUG(llvm::dbgs() << ">>>> Bottom Up!\n");
   SILBasicBlock &BB = BBState.getBB();
 
@@ -337,7 +514,7 @@ processBBBottomUp(ARCBBState &BBState,
       NestingDetected |= State.initWithInst(&I);
 
       DEBUG(llvm::dbgs() << "    REF COUNT DECREMENT! Known Safe: "
-            << (State.isKnownSafe()?"yes":"no") << "\n");
+                         << (State.isKnownSafe() ? "yes" : "no") << "\n");
 
       // Continue on to see if our reference decrement could potentially affect
       // any other pointers via a use or a decrement.
@@ -360,18 +537,18 @@ processBBBottomUp(ARCBBState &BBState,
         // Copy the current value of ref count state into the result map.
         IncToDecStateMap[&I] = RefCountState;
         DEBUG(llvm::dbgs() << "    MATCHING DECREMENT:"
-              << RefCountState.getValue());
+                           << RefCountState.getValue());
 
         // Clear the ref count state in case we see more operations on this
         // ref counted value. This is for safety reasons.
         RefCountState.clear();
       } else {
         if (RefCountState.isTrackingRefCountInst()) {
-          DEBUG(llvm::dbgs() << "    FAILED MATCH DECREMENT:"
-                << RefCountState.getValue());
+          DEBUG(llvm::dbgs()
+                << "    FAILED MATCH DECREMENT:" << RefCountState.getValue());
         } else {
           DEBUG(llvm::dbgs() << "    FAILED MATCH DECREMENT. Not tracking a "
-                "decrement.\n");
+                                "decrement.\n");
         }
       }
 
@@ -387,8 +564,8 @@ processBBBottomUp(ARCBBState &BBState,
       if (Op && OtherState.first == Op)
         continue;
 
-      // If we are tracking an argument, skip it.
-      if (!OtherState.second.isTrackingRefCountInst())
+      // If this state is not tracking anything, skip it.
+      if (!OtherState.second.isTrackingRefCount())
         continue;
 
       // Check if the instruction we are visiting could potentially decrement
@@ -396,7 +573,7 @@ processBBBottomUp(ARCBBState &BBState,
       // cause us to change states. If we do change states continue...
       if (OtherState.second.handlePotentialDecrement(&I, AA)) {
         DEBUG(llvm::dbgs() << "    Found Potential Decrement:\n        "
-              << OtherState.second.getValue());
+                           << OtherState.second.getValue());
         continue;
       }
 
@@ -404,7 +581,7 @@ processBBBottomUp(ARCBBState &BBState,
       // could be used by the given instruction.
       if (OtherState.second.handlePotentialUser(&I, AA))
         DEBUG(llvm::dbgs() << "    Found Potential Use:\n        "
-              << OtherState.second.getValue());
+                           << OtherState.second.getValue());
     }
   }
 
@@ -412,8 +589,8 @@ processBBBottomUp(ARCBBState &BBState,
 }
 
 void
-swift::arc::ARCSequenceDataflowEvaluator::
-mergeSuccessors(ARCBBState &BBState, SILBasicBlock *BB) {
+swift::arc::ARCSequenceDataflowEvaluator::mergeSuccessors(ARCBBState &BBState,
+                                                          SILBasicBlock *BB) {
   // Grab the backedge set for our BB.
   auto &BackEdgeSet = BackedgeMap[BB];
 
@@ -449,13 +626,18 @@ mergeSuccessors(ARCBBState &BBState, SILBasicBlock *BB) {
 bool swift::arc::ARCSequenceDataflowEvaluator::processBottomUp() {
   bool NestingDetected = false;
 
+  DEBUG(llvm::dbgs() << "<<<< Processing Bottom Up! >>>>\n");
+
   // For each BB in our post order...
   for (auto *BB : PostOrder) {
 
+    DEBUG(llvm::dbgs() << "Processing BB#: " << BBToPostOrderID[BB] << "\n");
+
     // Grab the BBState associated with it and set it to be the current BB.
     ARCBBState &BBState = BottomUpBBStates[BB];
     BBState.init(BB);
 
+    DEBUG(llvm::dbgs() << "Merging Successors!\n");
     mergeSuccessors(BBState, BB);
 
     // Then perform the basic block optimization.
@@ -509,8 +691,7 @@ bool swift::arc::ARCSequenceDataflowEvaluator::run() {
 //===----------------------------------------------------------------------===//
 
 bool swift::arc::performARCSequenceDataflow(
-    SILFunction &F,
-    AliasAnalysis *AA,
+    SILFunction &F, AliasAnalysis *AA,
     BlotMapVector<SILInstruction *, TopDownRefCountState> &DecToIncStateMap,
     BlotMapVector<SILInstruction *, BottomUpRefCountState> &IncToDecStateMap) {
 
diff --git a/lib/SILAnalysis/ReferenceCountState.h b/lib/SILAnalysis/ReferenceCountState.h
index db7434d5215..bb086f72901 100644
--- a/lib/SILAnalysis/ReferenceCountState.h
+++ b/lib/SILAnalysis/ReferenceCountState.h
@@ -184,7 +184,7 @@ struct RefCountState {
   bool handlePotentialDecrement(SILInstruction *PotentialDecrement,
                                 AliasAnalysis *AA) {
     // If we are not tracking a ref count, just return false.
-    if (!isTrackingRefCountInst())
+    if (!isTrackingRefCount())
       return false;
 
     // If we can prove that Other can not use the pointer we are tracking,
@@ -282,13 +282,13 @@ struct BottomUpRefCountState : public RefCountState<BottomUpRefCountState> {
 
   /// Is this ref count initialized and tracking a ref count ptr.
   bool isTrackingRefCount() const {
-    return !Decrements.empty();
+    return Decrements.size();
   }
 
   /// Are we tracking an instruction currently? This returns false when given an
   /// uninitialized ReferenceCountState.
   bool isTrackingRefCountInst() const {
-    return !Decrements.empty();
+    return Decrements.size();
   }
 
   /// Are we tracking a source of ref counts? This currently means that we are
@@ -359,7 +359,7 @@ struct BottomUpRefCountState : public RefCountState<BottomUpRefCountState> {
     }
   }
 
-  void merge(const BottomUpRefCountState &Other);
+  bool merge(const BottomUpRefCountState &Other);
 };
 
 //===----------------------------------------------------------------------===//
@@ -507,7 +507,7 @@ struct TopDownRefCountState : public RefCountState<TopDownRefCountState> {
     }
   }
 
-  void merge(const TopDownRefCountState &Other);
+  bool merge(const TopDownRefCountState &Other);
 };
 
 } // end arc namespace
diff --git a/test/SILPasses/global-arc-opts-loops.sil.gyb b/test/SILPasses/global-arc-opts-loops.sil.gyb
new file mode 100644
index 00000000000..5c461892fd0
--- /dev/null
+++ b/test/SILPasses/global-arc-opts-loops.sil.gyb
@@ -0,0 +1,125 @@
+%# -*- mode: sil -*-
+// RUN: rm -rf %t
+// RUN: mkdir %t
+// RUN: %gyb %s > %t/global-arc-opts-loops.sil
+// RUN: sil-opt -global-arc-opts -verify %t/global-arc-opts-loops.sil | FileCheck %t/global-arc-opts-loops.sil
+
+%# Ignore the following admonition; it applies to the resulting .sil
+%# test file only.
+// DO NOT MODIFY THIS TEST FILE. IT IS AUTOMATICALLY GENERATED BY GYB.
+
+sil_stage canonical
+
+import Builtin
+
+// Utilities
+
+sil @user : $@thin (Builtin.NativeObject) -> ()
+
+// Loop form 1 consists of CFG of the following form:
+//
+// BB0 -> BB1
+// BB1 -> BB1
+// BB1 -> BB2
+//
+// Thus it is a simple loop without any interior basic blocks.
+//
+// We generate test retain release pairs at 
+% for i in range(16):
+
+// CHECK-LABEL: sil @loop_form_1_${i} : $@thin (Builtin.NativeObject) -> () {
+sil @loop_form_1_${i} : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+bb0(%0 : $Builtin.NativeObject):
+% if i & 1:
+// CHECK-NEXT: strong_retain
+  strong_retain %0 : $Builtin.NativeObject
+% end
+// CHECK-NEXT: br
+  br bb1
+
+// CHECK: bb1:
+bb1:
+% if (i & 2) and not (i & 4):
+// CHECK-NEXT: strong_retain
+% end
+% if i & 2:
+  strong_retain %0 : $Builtin.NativeObject
+% end
+% if not (i & 2) and (i & 4):
+// CHECK-NEXT: strong_release
+% end
+% if i & 4:
+  strong_release %0 : $Builtin.NativeObject
+% end
+// CHECK-NEXT: cond_br
+  cond_br undef, bb1, bb2
+
+// CHECK: bb2:
+bb2:
+% if i & 8:
+// CHECK-NEXT: strong_release
+  strong_release %0 : $Builtin.NativeObject
+% end
+%%1 = tuple ()
+  return %1 : $()
+}
+% end
+
+// Loop form 2 consists of CFGs of the following form:
+//
+// BB0 -> BB1
+// BB1 -> BB2
+// BB2 -> BB1
+// BB2 -> BB3
+//
+// I.e. a loop with a body of 2 and the jump into the loop not at the
+// latch.
+
+% for i in range(3):
+% for j in range(4):
+
+sil @loop_form_2_${i}_${j} : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0({{.*}}):
+bb0(%0 : $Builtin.NativeObject):
+% if i & 1:
+// CHECK-NEXT: strong_retain
+  strong_retain %0 : $Builtin.NativeObject
+% end
+  br bb1
+
+// CHECK: bb1:
+bb1:
+% if j == 0:
+// CHECK-NEXT: strong_retain
+  strong_retain %0 : $Builtin.NativeObject
+% end
+% if j == 1:
+// CHECK-NEXT: strong_release
+  strong_release %0 : $Builtin.NativeObject
+% end
+  br bb2
+
+// CHECK: bb2:
+bb2:
+% if j == 2:
+// CHECK-NEXT: strong_retain
+  strong_retain %0 : $Builtin.NativeObject
+% end
+% if j == 3:
+// CHECK-NEXT: strong_release
+  strong_release %0 : $Builtin.NativeObject
+% end
+  cond_br undef, bb1, bb3
+
+// CHECK: bb3:
+bb3:
+% if i & 2:
+// CHECK-NEXT: strong_release
+  strong_release %0 : $Builtin.NativeObject
+% end
+%%1 = tuple ()
+  return %1 : $()
+}
+
+% end
diff --git a/test/SILPasses/globalarcopts.sil b/test/SILPasses/globalarcopts.sil
index ccd3fc26714..9eb6d2a1036 100644
--- a/test/SILPasses/globalarcopts.sil
+++ b/test/SILPasses/globalarcopts.sil
@@ -4,6 +4,10 @@ sil_stage canonical
 
 import Builtin
 
+// Utilities
+
+sil @user : $@thin (Builtin.NativeObject) -> ()
+
 /////////////////
 // Basic Tests //
 /////////////////
@@ -94,8 +98,6 @@ bb0(%0 : $S, %1 : $Builtin.NativeObject):
   return %2 : $()
 }
 
-sil @user : $@thin (Builtin.NativeObject) -> ()
-
 // CHECK-LABEL: sil @dont_delete_retain_over_decrement_use : $@thin (Builtin.NativeObject) -> ()
 // CHECK: bb0
 // CHECK: strong_retain
@@ -310,8 +312,15 @@ bb0(%0 : $*Builtin.Int32):
 }
 
 // CHECK-LABEL: sil @silargument_retain_iterated : $@thin (@owned Builtin.NativeObject) -> ()
-// CHECK-NOT: strong_retain
-// CHECK-NOT: strong_release
+// CHECK: bb0
+// CHECK-NEXT: function_ref user
+// CHECK-NEXT: function_ref @user
+// CHECK-NEXT: apply
+// CHECK-NEXT: apply
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: apply
+// CHECK-NEXT: apply
+// CHECK-NEXT: strong_release
 sil @silargument_retain_iterated : $@thin (@owned Builtin.NativeObject) -> () {
 bb0(%0 : $Builtin.NativeObject):
   %1 = function_ref @user : $@thin (Builtin.NativeObject) -> ()
@@ -441,3 +450,572 @@ bb0(%0 : $RetainUser, %1 : $Builtin.NativeObject):
   strong_release %0 : $RetainUser
   return %1 : $Builtin.NativeObject
 }
+
+
+////////////////////
+// Multi-BB tests //
+////////////////////
+
+//////////////
+// Hammocks //
+//////////////
+
+// CHECK-LABEL: sil @hammock1 : $@thin (Builtin.NativeObject) -> () {
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @hammock1 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// This hammock can not be optimized.
+// CHECK-LABEL: sil @hammock2 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK: bb1:
+// CHECK-NEXT: strong_retain
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+sil @hammock2 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+/// This hammock can't be optimized.
+// CHECK-LABEL: sil @hammock3 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK: bb1:
+// CHECK-NEXT: strong_release
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+sil @hammock3 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// This should not be optimizable.
+// CHECK-LABEL: sil @hammock4 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+// CHECK: bb1:
+// CHECK-NEXT: strong_retain
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+sil @hammock4 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// CHECK-LABEL: sil @hammock5 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NOT: strong_release
+// CHECK-NOT: strong_retain
+// CHECK: bb1:
+// CHECK-NEXT: strong_release
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+sil @hammock5 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// CHECK-LABEL: sil @hammock6 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+// CHECK: bb1:
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+sil @hammock6 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb2
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// CHECK-LABEL: sil @hammock7 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK: bb1:
+// CHECK-NEXT: strong_retain
+// CHECK: bb2:
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @hammock7 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+// CHECK-LABEL: sil @hammock8 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NOT: strong_release
+// CHECK: bb1:
+// CHECK-NEXT: strong_release
+// CHECK-NOT: strong_retain
+// CHECK: bb2:
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @hammock8 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb2
+
+bb2:
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+////////////////////
+// Double Hammock //
+////////////////////
+
+// Make sure we do not do anything in the presense of double partial
+// applies. This is due to issues related to the two branches of the two
+// diamonds not being control dependent.
+// CHECK-LABEL: sil @double_hammock1 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: function_ref user
+// CHECK-NEXT: function_ref @user
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: apply
+// CHECK-NEXT: bb2
+// CHECK: bb2:
+// CHECK-NEXT: cond_br
+// CHECK: bb3:
+// CHECK-NEXT: apply
+// CHECK-NEXT: br
+// CHECK: bb4:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple
+// CHECK-NEXT: return
+sil @double_hammock1 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  %1 = function_ref @user : $@thin (Builtin.NativeObject) -> ()
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  apply %1(%0) : $@thin (Builtin.NativeObject) -> ()
+  br bb2
+
+bb2:
+  cond_br undef, bb3, bb4
+
+bb3:
+  apply %1(%0) : $@thin (Builtin.NativeObject) -> ()
+  br bb4
+
+bb4:
+  strong_release %0 : $Builtin.NativeObject
+  %9999 = tuple()
+  return %9999 : $()
+}
+
+//////////////
+// Diamonds //
+//////////////
+
+// CHECK-LABEL: sil @diamond1 : $@thin (Builtin.NativeObject) -> () {
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @diamond1 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb3
+
+bb2:
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond2 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond2 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb3
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond3 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond3 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  br bb3
+
+bb3:
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond4 : $@thin (Builtin.NativeObject) -> () {
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @diamond4 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond5 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond5 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond6 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond6 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb3
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond7 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond7 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  strong_release %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond8 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond8 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+/// CHECK-LABEL: sil @diamond9 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond9 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb3
+
+bb2:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+// CHECK-LABEL: sil @diamond10 : $@thin (Builtin.NativeObject) -> () {
+// CHECK-NOT: strong_retain
+// CHECK-NOT: strong_release
+sil @diamond10 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+/// CHECK-LABEL: sil @diamond11 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond11 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb2:
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}
+
+/// CHECK-LABEL: sil @diamond12 : $@thin (Builtin.NativeObject) -> () {
+// CHECK: bb0
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: cond_br
+// CHECK: bb1:
+// CHECK-NEXT: br
+// CHECK: bb2:
+// CHECK-NEXT: strong_retain
+// CHECK-NEXT: br
+// CHECK: bb3:
+// CHECK-NEXT: strong_release
+// CHECK-NEXT: tuple ()
+// CHECK-NEXT: return
+sil @diamond12 : $@thin (Builtin.NativeObject) -> () {
+bb0(%0 : $Builtin.NativeObject):
+  strong_retain %0 : $Builtin.NativeObject
+  cond_br undef, bb1, bb2
+
+bb1:
+  br bb3
+
+bb2:
+  strong_retain %0 : $Builtin.NativeObject
+  br bb3
+
+bb3:
+  strong_release %0 : $Builtin.NativeObject
+  %1 = tuple()
+  return %1 : $()
+}