diff --git a/lib/Transforms/Scalar/Reassociate.cpp b/lib/Transforms/Scalar/Reassociate.cpp
index fc3b38ee53b..a3c241dc0f2 100644
--- a/lib/Transforms/Scalar/Reassociate.cpp
+++ b/lib/Transforms/Scalar/Reassociate.cpp
@@ -1195,9 +1195,6 @@ bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
   if (X != Opnd2->getSymbolicPart())
     return false;
 
-  const APInt &C1 = Opnd1->getConstPart();
-  const APInt &C2 = Opnd2->getConstPart();
-
   // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
   int DeadInstNum = 1;
   if (Opnd1->getValue()->hasOneUse())
@@ -1215,6 +1212,8 @@ bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
     if (Opnd2->isOrExpr())
       std::swap(Opnd1, Opnd2);
 
+    const APInt &C1 = Opnd1->getConstPart();
+    const APInt &C2 = Opnd2->getConstPart();
     APInt C3((~C1) ^ C2);
 
     // Do not increase code size!
@@ -1230,6 +1229,8 @@ bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
   } else if (Opnd1->isOrExpr()) {
     // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
     //
+    const APInt &C1 = Opnd1->getConstPart();
+    const APInt &C2 = Opnd2->getConstPart();
     APInt C3 = C1 ^ C2;
     
     // Do not increase code size
@@ -1244,6 +1245,8 @@ bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
   } else {
     // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
     //
+    const APInt &C1 = Opnd1->getConstPart();
+    const APInt &C2 = Opnd2->getConstPart();
     APInt C3 = C1 ^ C2;
     Res = createAndInstr(I, X, C3);
   }
diff --git a/test/Transforms/Reassociate/xor_reassoc.ll b/test/Transforms/Reassociate/xor_reassoc.ll
index d371a9b5b68..b9353c7f81f 100644
--- a/test/Transforms/Reassociate/xor_reassoc.ll
+++ b/test/Transforms/Reassociate/xor_reassoc.ll
@@ -164,3 +164,30 @@ define void @xor_bug1() {
   %3 = and i64 undef, %2
   ret void
 }
+
+; The bug was that when the compiler optimize "(x | c1)" ^ "(x & c2)", it may
+; swap the two xor-subexpressions if they are not in canoninical order; however,
+; when optimizer swaps two sub-expressions, if forgot to swap the cached value
+; of c1 and c2 accordingly, hence cause the problem.
+;
+define i32 @xor_bug2(i32, i32, i32, i32) {
+  %5 = mul i32 %0, 123
+  %6 = add i32 %2, 24
+  %7 = add i32 %1, 8
+  %8 = and i32 %1, 3456789
+  %9 = or i32 %8,  4567890
+  %10 = and i32 %1, 543210987
+  %11 = or i32 %1, 891034567
+  %12 = and i32 %2, 255
+  %13 = xor i32 %9, %10
+  %14 = xor i32 %11, %13
+  %15 = xor i32 %5, %14
+  %16 = and i32 %3, 255
+  %17 = xor i32 %16, 42
+  %18 = add i32 %6, %7
+  %19 = add i32 %18, %12
+  %20 = add i32 %19, %15
+  ret i32 %20
+;CHECK: @xor_bug2
+;CHECK: xor i32 %5, 891034567
+}