diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -10152,7 +10152,10 @@
     return SDValue();
 
   LoadSDNode *LN0 = cast<LoadSDNode>(N0);
-  if (!isLegalNarrowLdSt(LN0, ExtType, ExtVT, ShAmt))
+  // Reducing the width of a volatile load is illegal.  For atomics, we may be
+  // able to reduce the width provided we never widen again. (see D66309)  
+  if (!LN0->isSimple() ||
+      !isLegalNarrowLdSt(LN0, ExtType, ExtVT, ShAmt))
     return SDValue();
 
   auto AdjustBigEndianShift = [&](unsigned ShAmt) {
@@ -16276,6 +16279,11 @@
   if (OptLevel == CodeGenOpt::None)
     return SDValue();
 
+  // Can't change the number of memory accesses for a volatile store or break
+  // atomicity for an atomic one.
+  if (!ST->isSimple())
+    return SDValue();
+
   SDValue Val = ST->getValue();
   SDLoc DL(ST);
 
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -4859,6 +4859,8 @@
 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
                                               ISD::LoadExtType ExtTy,
                                               EVT NewVT) const {
+  assert(cast<LoadSDNode>(Load)->isSimple() && "illegal to narrow");
+  
   // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
   // relocation target a movq or addq instruction: don't let the load shrink.
   SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
@@ -7724,7 +7726,7 @@
 static bool findEltLoadSrc(SDValue Elt, LoadSDNode *&Ld, int64_t &ByteOffset) {
   if (ISD::isNON_EXTLoad(Elt.getNode())) {
     auto *BaseLd = cast<LoadSDNode>(Elt);
-    if (BaseLd->getMemOperand()->getFlags() & MachineMemOperand::MOVolatile)
+    if (!BaseLd->isSimple())
       return false;
     Ld = BaseLd;
     ByteOffset = 0;
@@ -7878,8 +7880,8 @@
 
   auto CreateLoad = [&DAG, &DL, &Loads](EVT VT, LoadSDNode *LDBase) {
     auto MMOFlags = LDBase->getMemOperand()->getFlags();
-    assert(!(MMOFlags & MachineMemOperand::MOVolatile) &&
-           "Cannot merge volatile loads.");
+    assert(LDBase->isSimple() &&
+           "Cannot merge volatile or atomic loads.");
     SDValue NewLd =
         DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
                     LDBase->getPointerInfo(), LDBase->getAlignment(), MMOFlags);