diff --git a/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp b/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
index 46b8221..468c561 100644
--- a/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
+++ b/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
@@ -88,6 +88,8 @@
 
   bool SelectTSTBOperand(SDValue N, SDValue &FixedPos, unsigned RegWidth);
 
+  SDNode *SelectAtomic(SDNode *N, unsigned Op8, unsigned Op16, unsigned Op32, unsigned Op64);
+
   SDNode *TrySelectToMoveImm(SDNode *N);
   SDNode *LowerToFPLitPool(SDNode *Node);
   SDNode *SelectToLitPool(SDNode *N);
@@ -318,6 +320,38 @@
   return true;
 }
 
+SDNode *AArch64DAGToDAGISel::SelectAtomic(SDNode *Node, unsigned Op8,
+                                          unsigned Op16,unsigned Op32,
+                                          unsigned Op64) {
+  // Mostly direct translation to the given operations, except that we preserve
+  // the AtomicOrdering for use later on.
+  AtomicSDNode *AN = cast<AtomicSDNode>(Node);
+  EVT VT = AN->getMemoryVT();
+
+  unsigned Op;
+  if (VT == MVT::i8)
+    Op = Op8;
+  else if (VT == MVT::i16)
+    Op = Op16;
+  else if (VT == MVT::i32)
+    Op = Op32;
+  else if (VT == MVT::i64)
+    Op = Op64;
+  else
+    llvm_unreachable("Unexpected atomic operation");
+
+  SmallVector<SDValue, 4> Ops;
+  for (unsigned i = 1; i < AN->getNumOperands(); ++i)
+      Ops.push_back(AN->getOperand(i));
+
+  Ops.push_back(CurDAG->getTargetConstant(AN->getOrdering(), MVT::i32));
+  Ops.push_back(AN->getOperand(0)); // Chain moves to the end
+
+  return CurDAG->SelectNodeTo(Node, Op,
+                              AN->getValueType(0), MVT::Other,
+                              &Ops[0], Ops.size());
+}
+
 SDNode *AArch64DAGToDAGISel::Select(SDNode *Node) {
   // Dump information about the Node being selected
   DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << "\n");
@@ -328,6 +362,78 @@
   }
 
   switch (Node->getOpcode()) {
+  case ISD::ATOMIC_LOAD_ADD:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_ADD_I8,
+                        AArch64::ATOMIC_LOAD_ADD_I16,
+                        AArch64::ATOMIC_LOAD_ADD_I32,
+                        AArch64::ATOMIC_LOAD_ADD_I64);
+  case ISD::ATOMIC_LOAD_SUB:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_SUB_I8,
+                        AArch64::ATOMIC_LOAD_SUB_I16,
+                        AArch64::ATOMIC_LOAD_SUB_I32,
+                        AArch64::ATOMIC_LOAD_SUB_I64);
+  case ISD::ATOMIC_LOAD_AND:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_AND_I8,
+                        AArch64::ATOMIC_LOAD_AND_I16,
+                        AArch64::ATOMIC_LOAD_AND_I32,
+                        AArch64::ATOMIC_LOAD_AND_I64);
+  case ISD::ATOMIC_LOAD_OR:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_OR_I8,
+                        AArch64::ATOMIC_LOAD_OR_I16,
+                        AArch64::ATOMIC_LOAD_OR_I32,
+                        AArch64::ATOMIC_LOAD_OR_I64);
+  case ISD::ATOMIC_LOAD_XOR:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_XOR_I8,
+                        AArch64::ATOMIC_LOAD_XOR_I16,
+                        AArch64::ATOMIC_LOAD_XOR_I32,
+                        AArch64::ATOMIC_LOAD_XOR_I64);
+  case ISD::ATOMIC_LOAD_NAND:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_NAND_I8,
+                        AArch64::ATOMIC_LOAD_NAND_I16,
+                        AArch64::ATOMIC_LOAD_NAND_I32,
+                        AArch64::ATOMIC_LOAD_NAND_I64);
+  case ISD::ATOMIC_LOAD_MIN:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_MIN_I8,
+                        AArch64::ATOMIC_LOAD_MIN_I16,
+                        AArch64::ATOMIC_LOAD_MIN_I32,
+                        AArch64::ATOMIC_LOAD_MIN_I64);
+  case ISD::ATOMIC_LOAD_MAX:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_MAX_I8,
+                        AArch64::ATOMIC_LOAD_MAX_I16,
+                        AArch64::ATOMIC_LOAD_MAX_I32,
+                        AArch64::ATOMIC_LOAD_MAX_I64);
+  case ISD::ATOMIC_LOAD_UMIN:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_UMIN_I8,
+                        AArch64::ATOMIC_LOAD_UMIN_I16,
+                        AArch64::ATOMIC_LOAD_UMIN_I32,
+                        AArch64::ATOMIC_LOAD_UMIN_I64);
+  case ISD::ATOMIC_LOAD_UMAX:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_LOAD_UMAX_I8,
+                        AArch64::ATOMIC_LOAD_UMAX_I16,
+                        AArch64::ATOMIC_LOAD_UMAX_I32,
+                        AArch64::ATOMIC_LOAD_UMAX_I64);
+  case ISD::ATOMIC_SWAP:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_SWAP_I8,
+                        AArch64::ATOMIC_SWAP_I16,
+                        AArch64::ATOMIC_SWAP_I32,
+                        AArch64::ATOMIC_SWAP_I64);
+  case ISD::ATOMIC_CMP_SWAP:
+    return SelectAtomic(Node,
+                        AArch64::ATOMIC_CMP_SWAP_I8,
+                        AArch64::ATOMIC_CMP_SWAP_I16,
+                        AArch64::ATOMIC_CMP_SWAP_I32,
+                        AArch64::ATOMIC_CMP_SWAP_I64);
   case ISD::FrameIndex: {
     int FI = cast<FrameIndexSDNode>(Node)->getIndex();
     EVT PtrTy = TLI.getPointerTy();
diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index e9f4497..6deae75 100644
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -59,12 +59,9 @@
 
   computeRegisterProperties();
 
-  // Some atomic operations can be folded into load-acquire or store-release
-  // instructions on AArch64. It's marginally simpler to let LLVM expand
-  // everything out to a barrier and then recombine the (few) barriers we can.
-  setInsertFencesForAtomic(true);
-  setTargetDAGCombine(ISD::ATOMIC_FENCE);
-  setTargetDAGCombine(ISD::ATOMIC_STORE);
+  // We have particularly efficient implementations of atomic fences if they can
+  // be combined with nearby atomic loads and stores.
+  setShouldFoldAtomicFences(true);
 
   // We combine OR nodes for bitfield and NEON BSL operations.
   setTargetDAGCombine(ISD::OR);
@@ -275,27 +272,34 @@
   return VT.changeVectorElementTypeToInteger();
 }
 
-static void getExclusiveOperation(unsigned Size, unsigned &ldrOpc,
-                                  unsigned &strOpc) {
-  switch (Size) {
-  default: llvm_unreachable("unsupported size for atomic binary op!");
-  case 1:
-    ldrOpc = AArch64::LDXR_byte;
-    strOpc = AArch64::STXR_byte;
-    break;
-  case 2:
-    ldrOpc = AArch64::LDXR_hword;
-    strOpc = AArch64::STXR_hword;
-    break;
-  case 4:
-    ldrOpc = AArch64::LDXR_word;
-    strOpc = AArch64::STXR_word;
-    break;
-  case 8:
-    ldrOpc = AArch64::LDXR_dword;
-    strOpc = AArch64::STXR_dword;
-    break;
-  }
+static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
+                                  unsigned &LdrOpc,
+                                  unsigned &StrOpc) {
+  static unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
+                                 AArch64::LDXR_word, AArch64::LDXR_dword};
+  static unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
+                                AArch64::LDAXR_word, AArch64::LDAXR_dword};
+  static unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
+                                  AArch64::STXR_word, AArch64::STXR_dword};
+  static unsigned StoreRels[] = {AArch64::STLXR_byte, AArch64::STLXR_hword,
+                                 AArch64::STLXR_word, AArch64::STLXR_dword};
+
+  unsigned *LoadOps, *StoreOps;
+  if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
+    LoadOps = LoadAcqs;
+  else
+    LoadOps = LoadBares;
+
+  if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
+    StoreOps = StoreRels;
+  else
+    StoreOps = StoreBares;
+
+  assert(isPowerOf2_32(Size) && Size <= 8 &&
+         "unsupported size for atomic binary op!");
+
+  LdrOpc = LoadOps[Log2_32(Size)];
+  StrOpc = StoreOps[Log2_32(Size)];
 }
 
 MachineBasicBlock *
@@ -313,12 +317,13 @@
   unsigned dest = MI->getOperand(0).getReg();
   unsigned ptr = MI->getOperand(1).getReg();
   unsigned incr = MI->getOperand(2).getReg();
+  AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
   DebugLoc dl = MI->getDebugLoc();
 
   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
 
   unsigned ldrOpc, strOpc;
-  getExclusiveOperation(Size, ldrOpc, strOpc);
+  getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
 
   MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
@@ -397,6 +402,8 @@
   unsigned dest = MI->getOperand(0).getReg();
   unsigned ptr = MI->getOperand(1).getReg();
   unsigned incr = MI->getOperand(2).getReg();
+  AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
+
   unsigned oldval = dest;
   DebugLoc dl = MI->getDebugLoc();
 
@@ -411,7 +418,7 @@
   }
 
   unsigned ldrOpc, strOpc;
-  getExclusiveOperation(Size, ldrOpc, strOpc);
+  getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
 
   MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
@@ -479,6 +486,7 @@
   unsigned ptr     = MI->getOperand(1).getReg();
   unsigned oldval  = MI->getOperand(2).getReg();
   unsigned newval  = MI->getOperand(3).getReg();
+  AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
   const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
   DebugLoc dl = MI->getDebugLoc();
 
@@ -487,7 +495,7 @@
   TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
 
   unsigned ldrOpc, strOpc;
-  getExclusiveOperation(Size, ldrOpc, strOpc);
+  getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
 
   MachineFunction *MF = BB->getParent();
   const BasicBlock *LLVM_BB = BB->getBasicBlock();
@@ -2377,78 +2385,6 @@
                      DAG.getConstant(LSB + Width - 1, MVT::i64));
 }
 
-static SDValue PerformATOMIC_FENCECombine(SDNode *FenceNode,
-                                         TargetLowering::DAGCombinerInfo &DCI) {
-  // An atomic operation followed by an acquiring atomic fence can be reduced to
-  // an acquiring load. The atomic operation provides a convenient pointer to
-  // load from. If the original operation was a load anyway we can actually
-  // combine the two operations into an acquiring load.
-  SelectionDAG &DAG = DCI.DAG;
-  SDValue AtomicOp = FenceNode->getOperand(0);
-  AtomicSDNode *AtomicNode = dyn_cast<AtomicSDNode>(AtomicOp);
-
-  // A fence on its own can't be optimised
-  if (!AtomicNode)
-    return SDValue();
-
-  AtomicOrdering FenceOrder
-    = static_cast<AtomicOrdering>(FenceNode->getConstantOperandVal(1));
-  SynchronizationScope FenceScope
-    = static_cast<SynchronizationScope>(FenceNode->getConstantOperandVal(2));
-
-  if (FenceOrder != Acquire || FenceScope != AtomicNode->getSynchScope())
-    return SDValue();
-
-  // If the original operation was an ATOMIC_LOAD then we'll be replacing it, so
-  // the chain we use should be its input, otherwise we'll put our store after
-  // it so we use its output chain.
-  SDValue Chain = AtomicNode->getOpcode() == ISD::ATOMIC_LOAD ?
-    AtomicNode->getChain() : AtomicOp;
-
-  // We have an acquire fence with a handy atomic operation nearby, we can
-  // convert the fence into a load-acquire, discarding the result.
-  DebugLoc DL = FenceNode->getDebugLoc();
-  SDValue Op = DAG.getAtomic(ISD::ATOMIC_LOAD, DL, AtomicNode->getMemoryVT(),
-                             AtomicNode->getValueType(0),
-                             Chain,                  // Chain
-                             AtomicOp.getOperand(1), // Pointer
-                             AtomicNode->getMemOperand(), Acquire,
-                             FenceScope);
-
-  if (AtomicNode->getOpcode() == ISD::ATOMIC_LOAD)
-    DAG.ReplaceAllUsesWith(AtomicNode, Op.getNode());
-
-  return Op.getValue(1);
-}
-
-static SDValue PerformATOMIC_STORECombine(SDNode *N,
-                                         TargetLowering::DAGCombinerInfo &DCI) {
-  // A releasing atomic fence followed by an atomic store can be combined into a
-  // single store operation.
-  SelectionDAG &DAG = DCI.DAG;
-  AtomicSDNode *AtomicNode = cast<AtomicSDNode>(N);
-  SDValue FenceOp = AtomicNode->getOperand(0);
-
-  if (FenceOp.getOpcode() != ISD::ATOMIC_FENCE)
-    return SDValue();
-
-  AtomicOrdering FenceOrder
-    = static_cast<AtomicOrdering>(FenceOp->getConstantOperandVal(1));
-  SynchronizationScope FenceScope
-    = static_cast<SynchronizationScope>(FenceOp->getConstantOperandVal(2));
-
-  if (FenceOrder != Release || FenceScope != AtomicNode->getSynchScope())
-    return SDValue();
-
-  DebugLoc DL = AtomicNode->getDebugLoc();
-  return DAG.getAtomic(ISD::ATOMIC_STORE, DL, AtomicNode->getMemoryVT(),
-                       FenceOp.getOperand(0),  // Chain
-                       AtomicNode->getOperand(1),       // Pointer
-                       AtomicNode->getOperand(2),       // Value
-                       AtomicNode->getMemOperand(), Release,
-                       FenceScope);
-}
-
 /// For a true bitfield insert, the bits getting into that contiguous mask
 /// should come from the low part of an existing value: they must be formed from
 /// a compatible SHL operation (unless they're already low). This function
@@ -2804,8 +2740,6 @@
   switch (N->getOpcode()) {
   default: break;
   case ISD::AND: return PerformANDCombine(N, DCI);
-  case ISD::ATOMIC_FENCE: return PerformATOMIC_FENCECombine(N, DCI);
-  case ISD::ATOMIC_STORE: return PerformATOMIC_STORECombine(N, DCI);
   case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
   case ISD::SRA: return PerformSRACombine(N, DCI);
   }
diff --git a/lib/Target/AArch64/AArch64InstrInfo.td b/lib/Target/AArch64/AArch64InstrInfo.td
index 37be5e4..ff21c22 100644
--- a/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/lib/Target/AArch64/AArch64InstrInfo.td
@@ -159,49 +159,55 @@
 // Atomic operation pseudo-instructions
 //===----------------------------------------------------------------------===//
 
-let usesCustomInserter = 1 in {
-multiclass AtomicSizes<string opname> {
-  def _I8 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
-          [(set i32:$dst, (!cast<SDNode>(opname # "_8") i64:$ptr, i32:$incr))]>;
-  def _I16 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
-         [(set i32:$dst, (!cast<SDNode>(opname # "_16") i64:$ptr, i32:$incr))]>;
-  def _I32 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
-         [(set i32:$dst, (!cast<SDNode>(opname # "_32") i64:$ptr, i32:$incr))]>;
-  def _I64 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$incr),
-         [(set i64:$dst, (!cast<SDNode>(opname # "_64") i64:$ptr, i64:$incr))]>;
+// These get selected from C++ code as a pretty much direct translation from the
+// generic DAG nodes. The one exception is the AtomicOrdering is added as an
+// operand so that the eventual lowering can make use of it and choose
+// acquire/release operations when required.
+
+let usesCustomInserter = 1, hasCtrlDep = 1, mayLoad = 1, mayStore = 1 in {
+multiclass AtomicSizes {
+  def _I8 : PseudoInst<(outs GPR32:$dst),
+                       (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+  def _I16 : PseudoInst<(outs GPR32:$dst),
+                        (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+  def _I32 : PseudoInst<(outs GPR32:$dst),
+                        (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+  def _I64 : PseudoInst<(outs GPR64:$dst),
+                        (ins GPR64xsp:$ptr, GPR64:$incr, i32imm:$ordering), []>;
 }
 }
 
-defm ATOMIC_LOAD_ADD  : AtomicSizes<"atomic_load_add">;
-defm ATOMIC_LOAD_SUB  : AtomicSizes<"atomic_load_sub">;
-defm ATOMIC_LOAD_AND  : AtomicSizes<"atomic_load_and">;
-defm ATOMIC_LOAD_OR   : AtomicSizes<"atomic_load_or">;
-defm ATOMIC_LOAD_XOR  : AtomicSizes<"atomic_load_xor">;
-defm ATOMIC_LOAD_NAND : AtomicSizes<"atomic_load_nand">;
-defm ATOMIC_SWAP      : AtomicSizes<"atomic_swap">;
+defm ATOMIC_LOAD_ADD  : AtomicSizes;
+defm ATOMIC_LOAD_SUB  : AtomicSizes;
+defm ATOMIC_LOAD_AND  : AtomicSizes;
+defm ATOMIC_LOAD_OR   : AtomicSizes;
+defm ATOMIC_LOAD_XOR  : AtomicSizes;
+defm ATOMIC_LOAD_NAND : AtomicSizes;
+defm ATOMIC_SWAP      : AtomicSizes;
 let Defs = [NZCV] in {
   // These operations need a CMP to calculate the correct value
-  defm ATOMIC_LOAD_MIN  : AtomicSizes<"atomic_load_min">;
-  defm ATOMIC_LOAD_MAX  : AtomicSizes<"atomic_load_max">;
-  defm ATOMIC_LOAD_UMIN : AtomicSizes<"atomic_load_umin">;
-  defm ATOMIC_LOAD_UMAX : AtomicSizes<"atomic_load_umax">;
+  defm ATOMIC_LOAD_MIN  : AtomicSizes;
+  defm ATOMIC_LOAD_MAX  : AtomicSizes;
+  defm ATOMIC_LOAD_UMIN : AtomicSizes;
+  defm ATOMIC_LOAD_UMAX : AtomicSizes;
 }
 
-let usesCustomInserter = 1, Defs = [NZCV] in {
-def ATOMIC_CMP_SWAP_I8
-  : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
-            [(set i32:$dst, (atomic_cmp_swap_8 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I16
-  : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
-           [(set i32:$dst, (atomic_cmp_swap_16 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I32
-  : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
-           [(set i32:$dst, (atomic_cmp_swap_32 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I64
-  : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$old, GPR64:$new),
-           [(set i64:$dst, (atomic_cmp_swap_64 i64:$ptr, i64:$old, i64:$new))]>;
+class AtomicCmpSwap<RegisterClass GPRData>
+  : PseudoInst<(outs GPRData:$dst),
+               (ins GPR64xsp:$ptr, GPRData:$old, GPRData:$new,
+                    i32imm:$ordering), []> {
+  let usesCustomInserter = 1;
+  let hasCtrlDep = 1;
+  let mayLoad = 1;
+  let mayStore = 1;
+  let Defs = [NZCV];
 }
 
+def ATOMIC_CMP_SWAP_I8  : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I16 : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I32 : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I64 : AtomicCmpSwap<GPR64>;
+
 //===----------------------------------------------------------------------===//
 // Add-subtract (extended register) instructions
 //===----------------------------------------------------------------------===//
@@ -2579,7 +2585,8 @@
 
 class acquiring_load<PatFrag base>
   : PatFrag<(ops node:$ptr), (base node:$ptr), [{
-  return cast<AtomicSDNode>(N)->getOrdering() == Acquire;
+  AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
+  return Ordering == Acquire || Ordering == SequentiallyConsistent;
 }]>;
 
 def atomic_load_acquire_8  : acquiring_load<atomic_load_8>;
@@ -2610,7 +2617,8 @@
 
 class releasing_store<PatFrag base>
   : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
-  return cast<AtomicSDNode>(N)->getOrdering() == Release;
+  AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
+  return Ordering == Release || Ordering == SequentiallyConsistent;
 }]>;
 
 def atomic_store_release_8  : releasing_store<atomic_store_8>;
diff --git a/test/CodeGen/AArch64/atomic-ops-not-barriers.ll b/test/CodeGen/AArch64/atomic-ops-not-barriers.ll
index 3c03e47..9888a74 100644
--- a/test/CodeGen/AArch64/atomic-ops-not-barriers.ll
+++ b/test/CodeGen/AArch64/atomic-ops-not-barriers.ll
@@ -1,4 +1,4 @@
-; RUN: llc -mtriple=aarch64-none-linux-gnu < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-none-linux-gnu -verify-machineinstrs < %s | FileCheck %s
 
 define i32 @foo(i32* %var, i1 %cond) {
 ; CHECK: foo:
@@ -9,7 +9,9 @@
   store i32 %newval, i32* %var
   br label %somewhere
 atomic_ver:
-  %val = atomicrmw add i32* %var, i32 -1 seq_cst
+  fence seq_cst
+  %val = atomicrmw add i32* %var, i32 -1 monotonic
+  fence seq_cst
   br label %somewhere
 ; CHECK: dmb
 ; CHECK: ldxr
diff --git a/test/CodeGen/AArch64/atomic-ops.ll b/test/CodeGen/AArch64/atomic-ops.ll
index f3c1617..5e87f21 100644
--- a/test/CodeGen/AArch64/atomic-ops.ll
+++ b/test/CodeGen/AArch64/atomic-ops.ll
@@ -8,18 +8,18 @@
 define i8 @test_atomic_load_add_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i8:
    %old = atomicrmw add i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -27,19 +27,19 @@
 
 define i16 @test_atomic_load_add_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i16:
-   %old = atomicrmw add i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -47,8 +47,8 @@
 
 define i32 @test_atomic_load_add_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i32:
-   %old = atomicrmw add i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -57,9 +57,9 @@
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -67,8 +67,8 @@
 
 define i64 @test_atomic_load_add_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i64:
-   %old = atomicrmw add i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -79,7 +79,7 @@
 ; CHECK-NEXT: add [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -87,8 +87,8 @@
 
 define i8 @test_atomic_load_sub_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i8:
-   %old = atomicrmw sub i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -99,7 +99,7 @@
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -107,8 +107,8 @@
 
 define i16 @test_atomic_load_sub_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i16:
-   %old = atomicrmw sub i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -117,9 +117,9 @@
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -127,19 +127,19 @@
 
 define i32 @test_atomic_load_sub_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i32:
-   %old = atomicrmw sub i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i32* @var32, i32 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -148,18 +148,18 @@
 define i64 @test_atomic_load_sub_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i64:
    %old = atomicrmw sub i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: sub [[NEW:x[0-9]+]], x[[OLD]], x0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -167,8 +167,8 @@
 
 define i8 @test_atomic_load_and_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i8:
-   %old = atomicrmw and i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i8* @var8, i8 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -177,9 +177,9 @@
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -187,8 +187,8 @@
 
 define i16 @test_atomic_load_and_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i16:
-   %old = atomicrmw and i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -199,7 +199,7 @@
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -208,18 +208,18 @@
 define i32 @test_atomic_load_and_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i32:
    %old = atomicrmw and i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -227,19 +227,19 @@
 
 define i64 @test_atomic_load_and_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i64:
-   %old = atomicrmw and i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i64* @var64, i64 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: and [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -248,18 +248,18 @@
 define i8 @test_atomic_load_or_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i8:
    %old = atomicrmw or i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -267,8 +267,8 @@
 
 define i16 @test_atomic_load_or_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i16:
-   %old = atomicrmw or i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -279,7 +279,7 @@
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -287,19 +287,19 @@
 
 define i32 @test_atomic_load_or_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i32:
-   %old = atomicrmw or i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i32* @var32, i32 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -307,8 +307,8 @@
 
 define i64 @test_atomic_load_or_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i64:
-   %old = atomicrmw or i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i64* @var64, i64 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -317,9 +317,9 @@
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: orr [[NEW:x[0-9]+]], x[[OLD]], x0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -327,19 +327,19 @@
 
 define i8 @test_atomic_load_xor_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i8:
-   %old = atomicrmw xor i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i8* @var8, i8 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -347,8 +347,8 @@
 
 define i16 @test_atomic_load_xor_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i16:
-   %old = atomicrmw xor i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -357,9 +357,9 @@
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -368,18 +368,18 @@
 define i32 @test_atomic_load_xor_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i32:
    %old = atomicrmw xor i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -387,8 +387,8 @@
 
 define i64 @test_atomic_load_xor_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i64:
-   %old = atomicrmw xor i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -399,7 +399,7 @@
 ; CHECK-NEXT: eor [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -407,8 +407,8 @@
 
 define i8 @test_atomic_load_xchg_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i8:
-   %old = atomicrmw xchg i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -418,7 +418,7 @@
   ; function there.
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -427,17 +427,17 @@
 define i16 @test_atomic_load_xchg_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i16:
    %old = atomicrmw xchg i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -445,8 +445,8 @@
 
 define i32 @test_atomic_load_xchg_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i32:
-   %old = atomicrmw xchg i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -454,9 +454,9 @@
 ; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -464,18 +464,18 @@
 
 define i64 @test_atomic_load_xchg_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i64:
-   %old = atomicrmw xchg i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i64* @var64, i64 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], x0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -484,20 +484,20 @@
 
 define i8 @test_atomic_load_min_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i8:
-   %old = atomicrmw min i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i8* @var8, i8 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -505,8 +505,8 @@
 
 define i16 @test_atomic_load_min_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i16:
-   %old = atomicrmw min i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -516,9 +516,9 @@
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -526,8 +526,8 @@
 
 define i32 @test_atomic_load_min_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i32:
-   %old = atomicrmw min i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i32* @var32, i32 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -539,7 +539,7 @@
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -548,19 +548,19 @@
 define i64 @test_atomic_load_min_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i64:
    %old = atomicrmw min i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, gt
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -569,19 +569,19 @@
 define i8 @test_atomic_load_max_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i8:
    %old = atomicrmw max i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -589,20 +589,20 @@
 
 define i16 @test_atomic_load_max_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i16:
-   %old = atomicrmw max i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -610,8 +610,8 @@
 
 define i32 @test_atomic_load_max_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i32:
-   %old = atomicrmw max i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -621,9 +621,9 @@
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -631,8 +631,8 @@
 
 define i64 @test_atomic_load_max_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i64:
-   %old = atomicrmw max i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -644,7 +644,7 @@
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, lt
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -652,8 +652,8 @@
 
 define i8 @test_atomic_load_umin_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i8:
-   %old = atomicrmw umin i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -665,7 +665,7 @@
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -673,20 +673,20 @@
 
 define i16 @test_atomic_load_umin_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i16:
-   %old = atomicrmw umin i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], uxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -695,19 +695,19 @@
 define i32 @test_atomic_load_umin_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i32:
    %old = atomicrmw umin i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -715,20 +715,20 @@
 
 define i64 @test_atomic_load_umin_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i64:
-   %old = atomicrmw umin i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i64* @var64, i64 %offset acq_rel
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, hi
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -736,20 +736,20 @@
 
 define i8 @test_atomic_load_umax_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i8:
-   %old = atomicrmw umax i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i8* @var8, i8 %offset acq_rel
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], uxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -757,8 +757,8 @@
 
 define i16 @test_atomic_load_umax_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i16:
-   %old = atomicrmw umax i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -770,7 +770,7 @@
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -779,19 +779,19 @@
 define i32 @test_atomic_load_umax_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i32:
    %old = atomicrmw umax i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -799,8 +799,8 @@
 
 define i64 @test_atomic_load_umax_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i64:
-   %old = atomicrmw umax i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i64* @var64, i64 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -810,9 +810,9 @@
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, lo
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -820,13 +820,13 @@
 
 define i8 @test_atomic_cmpxchg_i8(i8 %wanted, i8 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i8:
-   %old = cmpxchg i8* @var8, i8 %wanted, i8 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i8* @var8, i8 %wanted, i8 %new acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: [[STARTAGAIN:.LBB[0-9]+_[0-9]+]]:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w[[OLD]], w0
@@ -834,7 +834,7 @@
   ; As above, w1 is a reasonable guess.
 ; CHECK: stxrb [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -843,20 +843,20 @@
 define i16 @test_atomic_cmpxchg_i16(i16 %wanted, i16 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i16:
    %old = cmpxchg i16* @var16, i16 %wanted, i16 %new seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: [[STARTAGAIN:.LBB[0-9]+_[0-9]+]]:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w[[OLD]], w0
 ; CHECK-NEXT: b.ne [[GET_OUT:.LBB[0-9]+_[0-9]+]]
   ; As above, w1 is a reasonable guess.
-; CHECK: stxrh [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
+; CHECK: stlxrh [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -864,8 +864,8 @@
 
 define i32 @test_atomic_cmpxchg_i32(i32 %wanted, i32 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i32:
-   %old = cmpxchg i32* @var32, i32 %wanted, i32 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i32* @var32, i32 %wanted, i32 %new release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -876,9 +876,9 @@
 ; CHECK-NEXT: cmp w[[OLD]], w0
 ; CHECK-NEXT: b.ne [[GET_OUT:.LBB[0-9]+_[0-9]+]]
   ; As above, w1 is a reasonable guess.
-; CHECK: stxr [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
+; CHECK: stlxr [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -886,8 +886,8 @@
 
 define i64 @test_atomic_cmpxchg_i64(i64 %wanted, i64 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i64:
-   %old = cmpxchg i64* @var64, i64 %wanted, i64 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i64* @var64, i64 %wanted, i64 %new monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -900,7 +900,7 @@
   ; As above, w1 is a reasonable guess.
 ; CHECK: stxr [[STATUS:w[0-9]+]], x1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -933,19 +933,26 @@
 define i8 @test_atomic_load_acquire_i8() nounwind {
 ; CHECK: test_atomic_load_acquire_i8:
   %val = load atomic i8* @var8 acquire, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
-
+; CHECK-NOT: dmb
 ; CHECK: ldarb w0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i8 %val
 }
 
 define i8 @test_atomic_load_seq_cst_i8() nounwind {
 ; CHECK: test_atomic_load_seq_cst_i8:
   %val = load atomic i8* @var8 seq_cst, align 1
-; CHECK: adrp x[[HIADDR:[0-9]+]], var8
-; CHECK: ldrb w0, [x[[HIADDR]], #:lo12:var8]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
+; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
+; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
+; CHECK: ldarb w0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i8 %val
 }
 
@@ -954,6 +961,7 @@
   %val = load atomic i16* @var16 monotonic, align 2
 ; CHECK-NOT: dmb
 ; CHECK: adrp x[[HIADDR:[0-9]+]], var16
+; CHECK-NOT: dmb
 ; CHECK: ldrh w0, [x[[HIADDR]], #:lo12:var16]
 ; CHECK-NOT: dmb
 
@@ -976,9 +984,13 @@
 define i64 @test_atomic_load_seq_cst_i64() nounwind {
 ; CHECK: test_atomic_load_seq_cst_i64:
   %val = load atomic i64* @var64 seq_cst, align 8
-; CHECK: adrp x[[HIADDR:[0-9]+]], var64
-; CHECK: ldr x0, [x[[HIADDR]], #:lo12:var64]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
+; CHECK: adrp [[HIADDR:x[0-9]+]], var64
+; CHECK-NOT: dmb
+; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var64
+; CHECK-NOT: dmb
+; CHECK: ldar x0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i64 %val
 }
 
@@ -1005,20 +1017,26 @@
 define void @test_atomic_store_release_i8(i8 %val) nounwind {
 ; CHECK: test_atomic_store_release_i8:
   store atomic i8 %val, i8* @var8 release, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
 ; CHECK: stlrb w0, [x[[ADDR]]]
-
+; CHECK-NOT: dmb
   ret void
 }
 
 define void @test_atomic_store_seq_cst_i8(i8 %val) nounwind {
 ; CHECK: test_atomic_store_seq_cst_i8:
   store atomic i8 %val, i8* @var8 seq_cst, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
 ; CHECK: stlrb w0, [x[[ADDR]]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
   ret void
 }
@@ -1026,9 +1044,11 @@
 define void @test_atomic_store_monotonic_i16(i16 %val) nounwind {
 ; CHECK: test_atomic_store_monotonic_i16:
   store atomic i16 %val, i16* @var16 monotonic, align 2
+; CHECK-NOT: dmb
 ; CHECK: adrp x[[HIADDR:[0-9]+]], var16
+; CHECK-NOT: dmb
 ; CHECK: strh w0, [x[[HIADDR]], #:lo12:var16]
-
+; CHECK-NOT: dmb
   ret void
 }
 
@@ -1039,7 +1059,9 @@
   %addr = inttoptr i64 %addr_int to i32*
 
   store atomic i32 %val, i32* %addr monotonic, align 4
+; CHECK-NOT: dmb
 ; CHECK: str w2, [x0, x1]
+; CHECK-NOT: dmb
 
   ret void
 }
@@ -1047,9 +1069,12 @@
 define void @test_atomic_store_release_i64(i64 %val) nounwind {
 ; CHECK: test_atomic_store_release_i64:
   store atomic i64 %val, i64* @var64 release, align 8
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var64
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var64
+; CHECK-NOT: dmb
 ; CHECK: stlr x0, [x[[ADDR]]]
-
+; CHECK-NOT: dmb
   ret void
 }