Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13005,22 +13005,6 @@
   LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
   if (LD->getBasePtr() != Ptr) return Result;  // Not from same pointer.
 
-  // The store should be chained directly to the load or be an operand of a
-  // tokenfactor.
-  if (LD == Chain.getNode())
-    ; // ok.
-  else if (Chain->getOpcode() != ISD::TokenFactor)
-    return Result; // Fail.
-  else {
-    bool isOk = false;
-    for (const SDValue &ChainOp : Chain->op_values())
-      if (ChainOp.getNode() == LD) {
-        isOk = true;
-        break;
-      }
-    if (!isOk) return Result;
-  }
-
   // This only handles simple types.
   if (V.getValueType() != MVT::i16 &&
       V.getValueType() != MVT::i32 &&
@@ -13057,6 +13041,36 @@
   // is aligned the same as the access width.
   if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
 
+  // For narrowing to be valid, it must be the case that the load the
+  // immediately preceeding memory operation before the store.
+  if (LD == Chain.getNode())
+    ; // ok.
+  else if (Chain->getOpcode() != ISD::TokenFactor)
+    return Result; // Fail.
+  else {
+    // If the chain is a token factor we can do so if LD is an operand _AND_
+    // none of the other Token Factor operands are successors to LD.
+    SmallPtrSet<const SDNode *, 8> Visited;
+    SmallVector<const SDNode *, 16> Worklist;
+
+    bool isOk = false;
+    for (const SDValue &ChainOp : Chain->op_values())
+      if (ChainOp.getNode() == LD)
+        isOk = true;
+      else
+        Worklist.push_back(ChainOp.getNode());
+    // Search non-LD operands's predecessors for LD. This is potentially
+    // expensive so we bound the search. To mitigate failures due to size of
+    // DAG, we also add LD's predecessors to the pruning list.
+
+    for (const SDValue &LdOp : LD->op_values())
+      Visited.insert(LdOp.getNode());
+
+    const unsigned int Max = 1024;
+    if (!isOk || SDNode::hasPredecessorHelper(LD, Visited, Worklist, Max))
+      return Result;
+  }
+
   Result.first = MaskedBytes;
   Result.second = NotMaskTZ/8;
   return Result;
Index: llvm/test/CodeGen/X86/pr37826.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/X86/pr37826.ll
@@ -0,0 +1,42 @@
+; RUN: llc -o - %s -mtriple=x86_64--unknown-linux-gnu | FileCheck %s
+
+; When compiled and run this should print zero. 
+
+@c = common local_unnamed_addr global i32 0, align 4
+@f = common local_unnamed_addr global i32 0, align 4
+@e = common local_unnamed_addr global i32 0, align 4
+@.str.1 = private unnamed_addr constant [4 x i8] c"%d\0A\00", align 1
+
+; We should only see a single store to f (a bytes store to f+3).
+define void @k(i32 %l) {
+; CHECK-LABEL: k:
+; CHECK-NOT: {{.*}}, f({{[^,]*}}
+; CHECK: movb {{.*}}, f+3(%rip)
+; CHECK-NOT: {{.*}}, f({{[^,]*}}
+; CHECK:    ret
+  %load = load i32, i32* @c, align 4
+  %load6 = load i32, i32* @f, align 4
+  %clear7 = and i32 %load6, 16777215
+  store i32 %clear7, i32* @c, align 4
+  %neg = and i32 %load6, 2097151
+  %value = xor i32 %neg, 2097151
+  store i32 %load, i32* @c, align 4
+  %t0 = load i32, i32* @e, align 4
+  %value15 = xor i32 %t0, %value
+  %clear16 = and i32 %load6, -16777216
+  %set17 = or i32 %value15, %clear16
+  store i32 %set17, i32* @f, align 4
+  %clear25 = and i32 %set17, -16777216
+  %set26 = or i32 %clear25, %clear7
+  store i32 %set26, i32* @f, align 4
+  ret void
+}
+
+declare i32 @printf(i8* nocapture readonly, ...)
+
+define i32 @main() {
+  tail call void @k(i32 1)
+  %load = load i32, i32* @f, align 4
+  %call = tail call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([4 x i8], [4 x i8]* @.str.1, i64 0, i64 0), i32 %load)
+  ret i32 0
+}